June 20, 2024

Post-Exploitation Activities on PAN-OS Devices: A Network-Based Analysis

This blog investigates the network-based activity detected by Darktrace in compromises stemming from the exploitation of a vulnerability in Palo Alto Networks firewall devices, namely CVE-2024-3400.

Written by

Adam Potter

Senior Cyber Analyst

Inside the SOC

Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.

Written by

Adam Potter

Senior Cyber Analyst

Jun 2024

Update:
Following the initial publication of this blog detailing exploitation campaigns utilizing the recently disclosed vulnerability, Darktrace analysts expanded the scope of the threat research investigation to identify potential earlier, pre-CVE disclosure, exploitation of CVE 2024-3400. While the majority of PAN-OS exploitation activity seen in the Darktrace customer base occurred after the public release of the CVE, Darktrace did also see tooling activity likely related to CVE-2024-3400 exploitation prior to the vulnerability's disclosure. Unlike the post-CVE-release exploitation activity, which largely reflected indiscriminate, opportunistic targeting of unpatched systems, these pre-CVE release activities likely represented selective targeting by more calculated actors.

Between March 26 and 28, Darktrace identified two Palo Alto firewall devices within the network of a public sector customer making HTTP GET requests utilizing both cURL and wget user agents, versions of which were seen in later compromise activity in April. The devices requested multiple shell script files (.sh) from rare external IP addresses. These IPs are likely associated with an operational relay box (ORB) network^[1]. The connections also occurred without a specified hostname lookup, suggesting the IPs were hardcoded into process code or already cached through unexpected running processes. One of the destination IPs was later confirmed by Palo Alto Network’s Unit 42 as associated with exploitation of the PAN-OS vulnerability^[2]. This observed activity closely resembles post-exploitation activity seen on affected firewall devices in mid-April. However, unlike the more disruptive and noisier follow-on exploitation activity seen in post-CVE-release incidents, the pre-CVE-release case observed by Darktrace appears to have been much more discreet, likely due to the relevant threat actor's desire to remain undetected.

--

Introduction

Perimeter devices such as firewalls, virtual private networks (VPNs), and intrusion prevention systems (IPS), have long been the target of adversarial actors attempting to gain access to internal networks. However, recent publications and public service announcements by leading public institutions underscore the increased emphasis threat actors are putting on leveraging such products to initiate compromises.

A blog post by the UK National Cyber Security Center (NCSC) released in early 2024 notes that as improvements are made in the detection of phishing email payloads, threat actors have again begun re-focusing efforts to exploiting network edge devices, many of which are not secure by design, as a means of breach initiation.^[i] As such, it comes as no surprise that new Common Vulnerabilities and Exposures (CVEs) are constantly discovered that exploit such internet-exposed systems.

Darktrace analysts frequently observe the impacts of such CVEs first through their investigations via Darktrace’s Security Operations Center (SOC). Beginning in April 2024, Darktrace’s SOC began handling alerts and customer requests for potential incidents involving Palo Alto Networks firewall devices. Just days prior, external researchers publicly disclosed what would later be classified as PAN-OS CVE-2024-3400, a form of remote command execution vulnerability that affects several versions of Palo Alto Networks’ firewall operating system (PAN-OS), namely PAN-OS 11.1, 11.0 and 10.2. At the time, multiple Darktrace customers were unaware of the recently announced vulnerability.

The increase in observed SOC activity for Palo Alto firewall devices, coupled with the public announcement of the new CVE prompted Darktrace researchers to look for evidence of PAN-OS exploitation on customer networks. Researchers also focused on documenting post-exploitation activity from threat actors leveraging the recently disclosed vulnerability.

As such, this blog highlights the network-based behaviors involved in the CVE-2024-3400 attack chains investigated by Darktrace’s SOC and Threat Research teams. Moreover, this investigation also provides a deeper insight into the post-compromise activities of threat actors leveraging the novel CVE. Such insights will not only prove relevant for cybersecurity teams looking to inhibit compromises in this specific instance, but also highlights general patterns of behavior by threat actors utilizing such CVEs to target internet-facing systems.

CVE-2024-3400

In mid-April 2024, the Darktrace SOC observed an uptick in activity involving recurring patterns of malicious activity from Palo Alto firewall appliances. In response to this trend, Darktrace initiated a Threat Research investigation into such activity to try and identify common factors and indicators across seemingly parallel events. Shortly before the Threat Research team opened their investigation, external researchers provided public details of CVE-2024-3400, a form of remote command execution vulnerability in the GlobalProtect feature on Palo Alto Network firewall devices running PAN-OS versions: 10.2, 11.0, and 11.1.^[ii]

In their proof of concept, security researchers at watchTowr demonstrated how an attacker can pass session ID (SESSID) values to these PAN-OS devices to request files that do not exist. In response, the system creates a zero-byte file with root privileges with the same name.^[iii] Log data is passed on devices running telemetry services to external servers through command line functionality.^[iv] Given this functionality, external actors could then request non-existent files in the SESSID containing command parameters which then be interpreted by the command line functionality.^[v] Although researchers first believed the exploit could only be used against devices running telemetry services, this was later discovered to be untrue.^[vi]

As details of CVE-2024-3400 began to surface, Darktrace’s Threat Research analysts quickly identified distinct overlaps in the observed activity on specific customer deployments and the post-exploitation behavior reported by external researchers. Given the parallels, Darktrace correlated the patterns of activity observed by the SOC team to exploitation of the newly discovered vulnerability in PAN-OS firewall appliances.

Campaign Analysis

Between the April and May 2024, Darktrace identified four main themes of post-exploitation activity involving Palo Alto Network firewall devices likely targeted via CVE-2024-3400: exploitation validation, shell command and tool retrieval, configuration data exfiltration, and ongoing command and control through encrypted channels and application protocols.

1. Exploit Validation and Further Vulnerability Enumeration

Many of the investigated attack chains began with malicious actors using out-of-band application security testing (OAST) services such as Interactsh to validate exploits against Palo Alto firewall appliances. This exploit validation activity typically resulted in devices attempting to contact unusual external endpoints (namely, subdomains of ‘oast[.]pro’, ‘oast[.]live’, ‘oast[.]site’, ‘oast[.]online’, ‘oast[.]fun’, ‘oast[.]me’, and ‘g3n[.]in’) associated with OAST services such as Interactsh. These services can be used by developers to inspect and debug internet traffic, but also have been easily abused by threat actors.

While attempted connections to OAST services do not alone indicate CVE-2024-3400 exploitation, the prevalence of such activities in observed Palo Alto firewall attack chains suggests widespread usage of these OAST services to validate initial access methods and possibly further enumerate systems for additional vulnerabilities.

Figure 1: Model alert log details showcasing a PAN-OS device making DNS queries for Interactsh domain names in what could be exploit validation, and/or further host enumeration.

2. Command and Payload Transmission

The most common feature across analyzed incidents was HTTP GET requests for shell scripts and Linux executable files (ELF) from external IPs associated with exploitation of the CVE. These HTTP requests were frequently initiated using the utilities, cURL and wget. On nearly every device likely targeted by threat actors leveraging the CVE, Darktrace analysts highlighted the retrieval of shell scripts that either featured enumeration commands, the removal of evidence of compromise activity, or commands to retrieve and start binaries on the destination device.

a) Shell Script Retrieval

Investigated devices commonly performed HTTP GET requests to retrieve shell command scripts. Despite this commonality, there was some degree of variety amongst the retrieved payloads and their affiliation with certain command tools. Several distinct types of shell commands and files were identified during the analyzed breaches. For example, some firewall devices were seen requesting .txt files associated with both Sliver C2, whose malicious use has previously been investigated by Darktrace, and Cobalt Strike. The target URIs of devices’ HTTP requests for these files included, “36shr.txt”, “2.txt”, “bin.txt”, and “data.txt”.

More interestingly, though, was the frequency with which analyzed systems requested bash scripts from rare external IP addresses, sometimes over non-standard ports for the HTTP protocol. These bash scripts would feature commands usually for the recipient system to check for certain existing files and or running processes. If the file did not exist, the system would then use cURL or wget to obtain content from external sites, change the permissions of the file, and then execute, sending output to dev/null as a means of likely defense evasion. In some scripts, the system would first make a new folder, and change directories prior to acquiring external content. Additionally, some samples highlighted multiple attempts at enumeration of the host system.

Figure 2: Packet capture (PCAP) data highlighting the incoming shell scripts providing instructions to use cURL to obtain external content, change the permissions of the file to execute, and then run the binary using the credentials and details provided.

Figure 3: PCAP data highlighting a variation of a shell script seen in an HTTP response processed by compromised devices. The script provides instructions to make a directory, retrieve and execute external content, and to hide the output.

Not every retrieved file that was not explicitly a binary featured bash scripts. Model alerts on some deployments also included file masquerading attempts by threat actors, whereby the Palo Alto firewall device would request content with a misleading extension in the URI. In one such instance, the requested URI, and HTTP response header suggests the returned content is an image/png, but the actual body response featured configuration parameters for a new daemon service to be run on the system.

Figure 4: PCAP data indicating configuration details likely for a new daemon on an investigated host. Such HTTP body content differs from the image/png extension within the request URI and declared content type in the HTTP response header.

Bash scripts analyzed across customer deployments also mirrored those identified by external security teams. External researchers previously reported on a series of identifiable shell commands in some cases of CVE-2024-3400 exploitation analyzed by their teams. Commands frequently involved a persistence mechanism they later labeled as the “UPSTYLE” backdoor.^[vii] This python-based program operates by reading commands hidden in error logs generated by 404 requests to the compromised server. The backdoor interprets the requests and writes the output to CSS files on the device. In many cases, Darktrace’s Threat Research team noted clear parallels between shell commands retrieved via HTTP GET request with those directly involving UPSTYLE. There were also matches with some URI patterns identified with the backdoor and requests observed on Darktrace deployments.

Figure 5: HTTP response data containing shell commands potentially relating to the UPSTYLE backdoor.

The presence of these UPSTYLE-related shell commands in response to Palo Alto firewall devices’ HTTP requests provides further evidence for initial exploitation of the CVE. Many bash scripts in examined cases interacted with folders and files likely related to CVE-2024-3400 exploitation. These scripts frequently sought to delete contents of certain folders, such as “/opt/panlogs/tmp/device_telemetry/minute/*” where evidence of exploitation would likely reside. Moreover, recursive removal and copy commands were frequently seen targeting CSS files within the GlobalProtect folder, already noted as the vulnerable element within PAN-OS versions. This evidence is further corroborated by host-based forensic analysis conducted by external researchers.^[viii]

Figure 6: PCAP data from investigated system indicating likely defense evasion by removing content on folders where CVE exploitation occurred.

b) Executable File Retrieval

Typically, following command processing, compromised Palo Alto firewall devices proceeded to make web requests for several unusual and potentially malicious files. Many such executables would be retrieved via processed scripts. While there a fair amount of variety in specific executables and binaries obtained, overall, these executables involved either further command tooling such as Sliver C2 or Cobalt Strike payloads, or unknown executables. Affected systems would also employ uncommon ports for HTTP connections, in a likely attempt to evade detection. Extensions featured within the URI, when visible, frequently noted ‘.elf’ (Linux executable) or ‘.exe’ payloads. While most derived hashes did not feature identifiable open-source intelligence (OSINT) details, some samples did have external information tying the sample to specific malware. For example, one such investigation featured a compromised system requesting a file with a hash identified as the Spark malware (backdoor) while another investigated case included a host requesting a known crypto-miner.

Figure 7: PCAP data highlighting compromised system retrieving ELF content from a rare external server running a simple Python HTTP server.

Figure 8: Darktrace model alert logs highlighting a device labeled “Palo Alto” making a HTTP request on an uncommon port for an executable file following likely CVE exploitation.

3. Configuration Data Exfiltration and Unusual HTTP POST Activity

During Darktrace’s investigations, there were also several instances of sensitive data exfiltration from PAN-OS firewall devices. Specifically, targeted systems were observed making HTTP POST requests via destination port 80 to rare external endpoints that OSINT sources associate with CVE-2024-3400 exploitation and activity. PCAP analysis of such HTTP requests revealed that they often contained sensitive configuration details of the targeted Palo Alto firewall devices, including the IP address, default gateway, domain, users, superusers, and password hashes, to name only a few. Threat actors frequently utilized Target URIs such as “/upload” in their HTTP POST requests of this multi-part boundary form data. Again, the User-Agent headers of these HTTP requests largely involved versions of cURL, typically 7.6.1, and wget.

Figure 9: PCAP datahighlighting Palo Alto Firewall device running vulnerable version of PAN-OSposting configuration details to rare external services via HTTP.

Figure 10: Model alert logs highlighting a Palo Alto firewall device performing HTTP POSTs to a rare external IP, without a prior hostname lookup, on an uncommon port using a URI associated with configuration data exfiltration across analyzed incidents

Figure 11: Examples of TargetURIs of HTTP POST requests involving base64 encoded IPs and potential dataegress.

4. Ongoing C2 and Miscellaneous Activity

Lastly, a smaller number of affected Palo Alto firewall devices were seen engaging in repeated beaconing and/or C2 communication via both encrypted and unencrypted protocols during and following the initial series of kill chain events. Such encrypted channels typically involved protocols such as TLS/SSL and SSH. This activity likely represented ongoing communication of targeted systems with attacker infrastructure. Model alerts typically highlighted unusual levels of repeated external connectivity to rare external IP addresses over varying lengths of time. In some investigated incidents, beaconing activity consisted of hundreds of thousands of connections over several days.

Figure 12: Advanced search details highlighting high levels of ongoing external communication to endpoints associated with C2 infrastructure exploiting CVE-2024-3400.

Some beaconing activity appears to have involved the use of the WebSocket protocol, as indicated by the appearance of “/ws” URIs and validated within packet captures. Such connections were then upgraded to an encrypted connection.

Figure 13: PCAP highlighting use of WebSocket protocol to engage in ongoing external connectivity to likely C2 infrastructure following CVE-2024-3400 compromise.

While not directly visible in all the deployments, some investigations also yielded evidence of attempts at further post-exploitation activity. For example, a handful of the analyzed binaries that were downloaded by examined devices had OSINT information suggesting a relation to crypto-mining malware strains. However, crypto-mining activity was not directly observed at this time. Furthermore, several devices also triggered model alerts relating to brute-forcing activity via several authentication protocols (namely, Keberos and RADIUS) during the time of compromise. This brute-force activity likely represented attempts to move laterally from the affected firewall system to deeper parts of the network.

Figure 14: Model alert logs noting repeated SSL connectivity to a Sliver C2-affiliated endpoint in what likely constitutes C2 connectivity.

Figure 15: Model alert logs featuring repeated RADIUS login failures from a compromised PAN-OS device using generic usernames, suggesting brute-force activity.

Conclusion

Between April and late May 2024, Darktrace’s SOC and Threat Research teams identified several instances of likely PAN-OS CVE-2024-3400 exploitation across the Darktrace customer base. The subsequent investigation yielded four major themes that categorize the observed network-based post-exploitation activity. These major themes were exploit validation activity, retrieval of binaries and shell scripts, data exfiltration via HTTP POST activity, and ongoing C2 communication with rare external endpoints. The insights shared in this article will hopefully contribute to the ongoing discussion within the cybersecurity community about how to handle the likely continued exploitation of this vulnerability. Moreover, this article may also help cybersecurity professionals better respond to future exploitation of not only Palo Alto PAN-OS firewall devices, but also of edge devices more broadly.

Threat actors will continue to discover and leverage new CVEs impacting edge infrastructure. Since it is not yet known which CVEs threat actors will exploit next, relying on rules and signatures for the detection of exploitation of such CVEs is not a viable approach. Darktrace’s anomaly-based approach to threat detection, however, is well positioned to robustly adapt to threat actors’ changing methods, since although threat actors can change the CVEs they exploit, they cannot change the fact that their exploitation of CVEs results in highly unusual patterns of activity.

Credit to Adam Potter, Cyber Analyst, Sam Lister, Senior Cyber Analyst

Appendices ‍

Pre-CVE-Release IoCs

38.54[.]104[.]14/3.sh
154.223[.]16[.]34/1.sh
154.223[.]16[.]34/co.sh
38.54[.]104[.]14/

Indicators of Compromise

Indicator – Type – Description

94.131.120[.]80 IP C2 Endpoint

94.131.120[.]80:53/?src=[REDACTED]=hour=root URL C2/Exfiltration Endpoint

134.213.29[.]14/?src=[REDACTED]min=root URL C2/Exfiltration Endpoint

134.213.29[.]14/grep[.]mips64 URL Payload

134.213.29[.]14/grep[.]x86_64 URL Payload

134.213.29[.]14/?deer URL Payload

134.213.29[.]14/?host=IDS URL Payload

134.213.29[.]14/ldr[.]sh URL Payload

91ebcea4e6d34fd6e22f99713eaf67571b51ab01 SHA1 File Hash Payload

185.243.115[.]250/snmpd2[.]elf URL Payload

23.163.0[.]111/com URL Payload

80.92.205[.]239/upload URL C2/Exfiltration Endpoint

194.36.171[.]43/upload URL C2/Exfiltration Endpoint

update.gl-protect[.]com Hostname C2 Endpoint

update.gl-protect[.]com:63869/snmpgp URL Payload

146.70.87[.]237 IP address C2 Endpoint

146.70.87[.]237:63867/snmpdd URL Payload

393c41b3ceab4beecf365285e8bdf0546f41efad SHA1 File Hash Payload

138.68.44[.]59/app/r URL Payload

138.68.44[.]59/app/clientr URL Payload

138.68.44[.]59/manage URL Payload

72.5.43[.]90/patch URL Payload

217.69.3[.]218 IP C2 Endpoint

5e8387c24b75c778c920f8aa38e4d3882cc6d306 SHA1 File Hash Payload

217.69.3[.]218/snmpd[.]elf URL Payload

958f13da6ccf98fcaa270a6e24f83b1a4832938a SHA1 File Hash Payload

6708dc41b15b892279af2947f143af95fb9efe6e SHA1 File Hash Payload

dc50c0de7f24baf03d4f4c6fdf6c366d2fcfbe6c SHA1 File Hash Payload

109.120.178[.]253:10000/data[.]txt URL Payload

109.120.178[.]253:10000/bin[.]txt URL Payload

bc9dc2e42654e2179210d98f77822723740a5ba6 SHA1 File Hash Payload

109.120.178[.]253:10000/123 URL Payload

65283921da4e8b5eabb926e60ca9ad3d087e67fa SHA1 File Hash Payload

img.dxyjg[.]com/6hiryXjZN0Mx[.]sh URL Payload

149.56.18[.]189/IC4nzNvf7w/2[.]txt URL Payload

228d05fd92ec4d19659d71693198564ae6f6b117 SHA1 File Hash Payload

54b892b8fdab7c07e1e123340d800e7ed0386600 SHA1 File Hash Payload

165.232.121[.]217/rules URL Payload

165.232.121[.]217/app/request URL Payload

938faec77ebdac758587bba999e470785253edaf SHA1 File Hash Payload

165.232.121[.]217/app/request63 URL Payload

165.232.121[.]217:4443/termite/165.232.121[.]217 URL Payload

92.118.112[.]60/snmpd2[.]elf URL Payload

2a90d481a7134d66e8b7886cdfe98d9c1264a386 SHA1 File Hash Payload

92.118.112[.]60/36shr[.]txt URL Payload

d6a33673cedb12811dde03a705e1302464d8227f SHA1 File Hash Payload

c712712a563fe09fa525dfc01ce13564e3d98d67 SHA1 File Hash Payload

091b3b33e0d1b55852167c3069afcdb0af5e5e79 SHA1 File Hash Payload

5eebf7518325e6d3a0fd7da2c53e7d229d7b74b6 SHA1 File Hash Payload

183be7a0c958f5ed4816c781a2d7d5aa8a0bca9f SHA1 File Hash Payload

e7d2f1224546b17d805617d02ade91a9a20e783e SHA1 File Hash Payload

e6137a15df66054e4c97e1f4b8181798985b480d SHA1 File Hash Payload

95.164.7[.]33:53/sea[.]png URL Payload

95.164.7[.]33/rules URL Payload

95.164.7[.]33:53/lb64 URL Payload

c2bc9a7657bea17792048902ccf2d77a2f50d2d7 SHA1 File Hash Payload

923369bbb86b9a9ccf42ba6f0d022b1cd4f33e9d SHA1 File Hash Payload

52972a971a05b842c6b90c581b5c697f740cb5b9 SHA1 File Hash Payload

95d45b455cf62186c272c03d6253fef65227f63a SHA1 File Hash Payload

322ec0942cef33b4c55e5e939407cd02e295973e SHA1 File Hash Payload

6335e08873b4ca3d0eac1ea265f89a9ef29023f2 SHA1 File Hash Payload

134.213.29[.]14 IP C2 Endpoint

185.243.115[.]250 IP C2 Endpoint

80.92.205[.]239 IP C2 Endpoint

194.36.171[.]43 IP C2 Endpoint

92.118.112[.]60 IP C2 Endpoint

109.120.178[.]253 IP C2 Endpoint

23.163.0[.]111 IP C2 Endpoint

72.5.43[.]90 IP C2 Endpoint

165.232.121[.]217 IP C2 Endpoint

8.210.242[.]112 IP C2 Endpoint

149.56.18[.]189 IP C2 Endpoint

95.164.7[.]33 IP C2 Endpoint

138.68.44[.]59 IP C2 Endpoint

Img[.]dxyjg[.]com Hostname C2 Endpoint

Darktrace Model Alert Coverage

· Anomalous Connection / New User Agent to IP Without Hostname

· Device / New User Agent (triggered by pre-CVE-release activity)

· Anomalous File / Script from Rare External Location (triggered by pre-CVE-release activity)

· Anomalous File / Masqueraded File Transfer

· Anomalous File / EXE from Rare External Location

· Anomalous File / Multiple EXE from Rare External Locations

· Anomalous File / Script and EXE from Rare External

· Anomalous File / Suspicious Octet Stream Download

· Anomalous File / Numeric File Download

· Anomalous Connection / Application Protocol on Uncommon Port

· Anomalous Connection / Posting HTTP to IP Without Hostname

· Anomalous Connection / Multiple Failed Connections to Rare Endpoint

· Anomalous Connection / Suspicious Self-Signed SSL

· Anomalous Connection / Anomalous SSL without SNI to New External

· Anomalous Connection / Multiple Connections to New External TCP Port

· Anomalous Connection / Rare External SSL Self-Signed

· Anomalous Server Activity / Outgoing from Server

· Anomalous Server Activity / Rare External from Server

· Compromise / SSH Beacon

· Compromise / Beacon for 4 Days

· Compromise / Sustained TCP Beaconing Activity To Rare Endpoint

· Compromise / High Priority Tunnelling to Bin Services

· Compromise / Sustained SSL or HTTP Increase

· Compromise / Connection to Suspicious SSL Server

· Compromise / Suspicious File and C2

· Compromise / Large Number of Suspicious Successful Connections

· Compromise / Slow Beaconing Activity To External Rare

· Compromise / HTTP Beaconing to New Endpoint

· Compromise / SSL or HTTP Beacon

· Compromise / Suspicious HTTP and Anomalous Activity

· Compromise / Beacon to Young Endpoint

· Compromise / High Volume of Connections with Beacon Score

· Compromise / Suspicious Beaconing Behaviour

· Compliance / SSH to Rare External Destination

· Compromise / HTTP Beaconing to Rare Destination

· Compromise / Beaconing Activity To External Rare

· Device / Initial Breach Chain Compromise

· Device / Multiple C2 Model Breaches

MITRE ATTACK Mapping

Tactic – Technique

Initial Access T1190 – Exploiting Public-Facing Application

Execution T1059.004 – Command and Scripting Interpreter: Unix Shell

Persistence T1543.002 – Create or Modify System Processes: Systemd Service

Defense Evasion T1070.004 – Indicator Removal: File Deletion

Credential Access T1110.001 – Brute Force: Password Guessing

Discovery T1083 – File and System Discovery

T1057 – Process Discovery

Collection T1005 – Data From Local System

Command and Control

T1071.001 – Application Layer Protocol: Web Protocols

T1573.002 – Encrypted Channel: Asymmetric Cryptography

T1571 – Non-Standard Port

T1105 – Ingress Tool Transfer

Exfiltration

T1041 – Exfiltration over C2 Protocol

T1048.002 - Exfiltration Over Alternative Protocol: Exfiltration Over Asymmetric Encrypted Non-C2 Protocol

References

[1] https://cloud.google.com/blog/topics/threat-intelligence/china-nexus-espionage-orb-networks

[2] https://unit42.paloaltonetworks.com/cve-2024-3400/

^[i] https://www.ncsc.gov.uk/blog-post/products-on-your-perimeter

^[ii] https://security.paloaltonetworks.com/CVE-2024-3400

^[iii] https://labs.watchtowr.com/palo-alto-putting-the-protecc-in-globalprotect-cve-2024-3400/

^[iv] https://labs.watchtowr.com/palo-alto-putting-the-protecc-in-globalprotect-cve-2024-3400/

^[v] https://labs.watchtowr.com/palo-alto-putting-the-protecc-in-globalprotect-cve-2024-3400/

^[vi] https://security.paloaltonetworks.com/CVE-2024-3400

^[vii] https://www.volexity.com/blog/2024/04/12/zero-day-exploitation-of-unauthenticated-remote-code-execution-vulnerability-in-globalprotect-cve-2024-3400/

^[viii] https://www.volexity.com/blog/2024/05/15/detecting-compromise-of-cve-2024-3400-on-palo-alto-networks-globalprotect-devices/

‍

Written by

Adam Potter

Senior Cyber Analyst

Inside the SOC

Written by

Adam Potter

Senior Cyber Analyst

Watch the NIS2 Webinar

Blog

July 8, 2026

Securing AI: Analysis of the Complete Security Stack with Governance and Controls

Why traditional cybersecurity approaches are not enough for AI

AI adoption outpaces most security programs’ ability to adapt. That gap is now one of the most consequential sources of cyber risk facing enterprises. As organizations embed generative and agentic AI into development workflows, business operations, and security tooling itself, the question is no longer whether AI will introduce risk. The question is whether organizations understand where that risk actually lives and how to manage it operationally.

Two recent pieces of guidance underscore this shift:

The upcoming Cybersecurity Framework Profile for AI from NIST

The Five Eyes government guidance on the careful adoption of agentic AI services

Taken together, they point to a critical conclusion. AI security cannot be reduced to model hardening or prompt filtering. It requires a defense in depth strategy that treats AI as both a new attack surface and a force multiplier for defense, while accounting for how AI fundamentally changes scale, speed, and autonomy.

Recent threat research suggests that today's cyber risk is driven less by initial compromise and more by an adversary's ability to blend into normal operations over time. AI systems create the same exposure in a new form: more autonomy, more scale, and more opportunities for risky behavior to blend into normal operations.

How NIST defines the three core pillars of AI security

The NIST profile organizes AI risk across three inseparable focus areas that span all cybersecurity functions, Secure, Defend and Thwart. These areas are not sequential. They exist simultaneously and must be addressed together.

Secure

This treats AI as an attack surface. It includes models, prompts, agents, pipelines, training and inference data, retrieval augmented generation corpora, and the AI supply chain itself. AI systems are opaque, probabilistic, and non-deterministic by design. Some vulnerabilities are inherent in how models are trained or how data is sourced. Traditional patching does not fully mitigate these risks. This is also where many enterprises are weakest today and, critically, where many security programs stop.

Defend

This is AI as a defensive force multiplier. AI can improve detection speed, scale, correlation, and response, but only if the right models are used and operationalized correctly. Machine-speed behavior-based detection, response and containment becomes critical in defending non-deterministic systems. Accuracy, explainability, governance, testing, validation, and integration into SOC workflows matter as much as capability. Without those controls, hallucination risk, over automation, and misplaced trust become security risks themselves.

Thwart

This treats AI as an adversarial accelerant. Threat actors are already using AI to generate targeted social engineering attacks, deepfakes, malware, and autonomous attack agents. Asymmetric warfare is highlighting faster vulnerability discovery and exploitation with a lag on patch development, testing and deployment.

How this looks in practice

Darktrace researchers observed scaled, automated exploitation of the React2Shell vulnerability within days of disclosure. A vulnerable cloud asset was exploited in under 120 seconds of being deployed. Darktrace research team observed an AI/LLM-generated malware sample used in exploitation activity tied to React2Shell. The significance isn't novelty. It is that AI lowers the barrier to producing usable offensive tooling and compresses the time between experimentation and deployment.

Tactics are getting more and more creative in order to string together steps of an attack kill chain. This creates a dependency on behavior-based detection, autonomous investigation, autonomous containment, training, resilience investment, and recovery planning across the entire enterprise.

Why agentic AI fundamentally changes enterprise cyber risk

The Five Eyes guidance on agentic AI highlights material changes to the cyber risk profile of an organization. Unlike generative AI systems that produce content for human consumption, agentic AI systems reason, plan, and act autonomously across tools, data, and environments. That autonomy, combined with access to real systems, amplifies the impact of traditional cyber failures and introduces new system level risks that are difficult to predict, observe, and contain.

Risk in agentic systems does not live in the model alone. It emerges from interactions between models, prompts, memory, tools, APIs, identities, privileges, inter-agent trust relationships, and human assumptions baked into design. Vulnerabilities are often introduced through data, connectors, natural language interfaces, protocols, and drift by design.

In supply-chain incidents, attackers did not need sophisticated exploits to scale impact. They abused trusted systems built for automation and implicit access. Agentic AI inherits that model. Once a system can act across tools, data, and workflows, compromise propagates through trust relationships that were never designed for machine autonomy.

The major agentic AI risk classes include the following:

The identity control for non-human identities or autonomous agents makes it difficult to mitigate over-permissioning, limiting access, scope, and duration, as well as access hygiene

Agents are frequently over permissioned

Compromised tools inherit agent authority

Static secrets enable impersonation

Implicit trust between agents enables lateral movement

Design and configuration risks compound this, including privileges evaluated once at startup, poor segmentation, unvetted third party tools, reused authorization decisions outside their original context, and guardrail limitations.

Behavioral risk

Agents can optimize for goals in unsafe ways, misinterpret ambiguous intent, chain actions into unintended sequences, change behavior during evaluation, and exhibit deceptive or sycophantic responses.

Structural risk

Structural risk follows from agentic systems that are tightly coupled, multicomponent ecosystems. Failures can propagate across agents. Hallucinations cascade downstream. Resource exhaustion becomes systemic. Tool misuse enables indirect prompt injection and command execution. Rogue agents can poison peer agents through trust relationships.

Accountability

Accountability becomes unclear as autonomy increases. Autonomous agents assume human identity permissions, and humans should have clear ownership of these agents, but they don’t, and this model is flawed. Decision paths are opaque and non-deterministic. Logs are fragmented and difficult to interpret. Reproducing an incident will be impossible without explicit design for observability and forensics. An agent compromise is functionally an insider threat, often with better access and fewer behavioral constraints than a human.

What does defense in depth look like for AI?

Agentic AI runs on software, networks, identities, and data. It must be governed using the same foundational principles that have proven resilient under uncertainty, including secure by design, defense in depth, zero trust, least privilege, continuous monitoring, behavior-based advanced threat detection and containment, and incident response and recovery.

Core components to a Defense in depth Strategy for Securing the use of AI:

Strong, precise identity control plane to include an identity per agent (cryptographic, non‑shared)
- Privilege monitoring and just‑in‑time access

Data Governance

Secure‑by‑default configurations
- Security Posture Management
- Zero Trust principles

Strong guardrails, deny‑by‑default policies, and isolation

Explicit instruction hierarchies and controlled context

Behavioral-based detection across entire enterprise to include inputs, tools, and outputs as well as AI used on the endpoint, across the network, cloud, SaaS, email, and OT
- Runtime anomaly detection and goal‑drift detection
- Autonomous containment to mitigate risk and minimize damage

Hard boundaries on autonomy and delegation

Testing, Evaluation, Validation and Verification
- Determine when autonomous action and when human in the loop
- Adversarial training and agent‑specific testing
- Simulation, red teaming, and chaos testing

Kill‑switches, rollback, and containment mechanisms
- Forensics data captures, interpretability, autonomous containment, and remediation/recovery plans

Until standards, tooling, and assurance methods mature, organizations should assume agentic AI systems will behave unexpectedly and design deployments around resilience, behavior-based detection, reversibility, and containment, not efficiency.

How security leaders should prepare for enterprise AI adoption

AI security is not model security alone. Data, pipelines, identities, and agents are first class assets. Many AI attacks succeed through standard cyber failures amplified by AI. Identity, data, and supply chain risk dominate. Behavior-based detection and response are critical, not optional. Logging, provenance, versioning, and forensics data capture of detections are mandatory because you cannot investigate or recover from AI incidents without them.

Risk will often be visible in behavior before it is clearly defined in policy or guidance. The same pattern has been seen in pre-CVE disclosure detection, where abnormal activity appears before the industry has named or described the vulnerability. AI systems introduce that uncertainty by design.

Security leaders should prioritize controls before AI is fully deployed, avoid generic AI security checklists, integrate AI risk into existing cyber programs, and mitigate the risk of non-deterministic technology with continuous oversight, monitoring, behavior analytics, anomaly detection, autonomous investigation, and autonomous containment.

Visibility has a different connotation with AI. Previously, audit logging worked for software/people, but with Generative AI-based systems, interpretability and explainability is difficult to understand, you cannot "undo" what has been done, or see the logic or control a chain of events. This is why behavioral-based detections and containment becomes critical.

What capabilities should every AI security program include?

If an organization asked “what must be in place before scaling AI?”:

AI Risk board and approval workflow

IAM + PAM for all AI services and agents

AI asset inventory

Prompt/output DLP with sanctioned AI access – This is not just pre- and post- filters, but behavior-based detections of semantic interface as well as behavior-based analysis of output with associated risk context.

Shadow AI identification

Secure MLOps – This is an entire paper itself

Runtime guardrails and tool restrictions
- Including AI Gateway/SASE/Zero trust/

Runtime security with behavior-based detections
- Complete visibility, monitoring, behavior analytics, anomaly detection, risk/intent/context evaluation of anomalies, autonomous investigation and autonomous containment of all AI assets across endpoint, network, SaaS, SASE, cloud, OT, email, and messaging platforms

Secure data pipelines and data governance

SOC workflow changes from malicious classification workflows to behavior-based detection workflows

Remediation plans for AI-related incidents

Layered Governance and Security Stack for Securing AI

The following outline considers governance and security tools that should be considered, well-integrated, deployed, tested, operationalized and embedded within security workflows. These tools and controls map to NIST’s CMF for AI.

These considerations do not need to be implemented in order. Runtime Detect and Respond will help mitigate risk while Governance, Visibility, and Identity mature.

Category	Tooling	Controls
Governance & Visibility	AI asset inventory / AI CMDB Shadow AI discovery SaaS discovery AI usage on non-endpoint managed systems via network or cloud telemetry MCP server/client usage via protocols Browser telemetry Gateway or SASE telemetry	Establish a risk board to set up controls Mandatory registration of AI systems Owner, data classification, intended use, and risk tier Supplier disclosure requirements Risk mitigation plan for AI adoption, innovation, or development
Identity, Access & Agent Control	Non-human autonomous agents should not have the full permissions associated with a human user. IAM with workload identities PAM for AI service accounts Secrets management with short-lived tokens Zero Trust principles Identity, permission, and token hygiene	Unique identities per model, agent, and pipeline Least privilege for tools, data, and APIs Explicit approval for autonomous actions
Data Security & Privacy	Data classification and labeling Enterprise DLP across endpoint, email, network, cloud, and SaaS Forensics data capture after risky detections	Prompt-level DLP through behavior-based semantic analysis with risk and intent context Input/interface analysis for risky data requests Output analysis for sensitive data Data integrity evaluation Retention and redaction policies for prompts and responses
Secure MLOps / LLMOps	Secure CI/CD with AI-specific gates Model registries with approval workflows Dependency, container, and artifact scanning SBOM/AIBOM generation IaC security scanning Security posture management Misconfiguration identification Hardening recommendations	Signed models and prompts Versioned datasets, configurations, logging, and controls Securing data pipelines Controlled promotion Quality assurance Adversarial testing
Runtime Security	Securing runtime goes beyond guardrails and model firewalls to include behavior-based detections, response, and containment. Detection, monitoring, and SOC integration Centralized visibility into prompts, outputs, and tool calls AI-specific detections Behavior-based detection for AI usage patterns Model drift and behavior monitoring Autonomous containment Behavior-based detection of model inputs and outputs Prompt injection detection Model manipulation, including jailbreaking, poisoning, and related attacks Sensitive data access attempts Behavior-based detection across low-code agents, high-code agents, MCP clients and servers, endpoint, network, cloud, email, SaaS, SASE, IoT, and OT	Policy enforcement between users, models, tools, agents, SaaS models/tools, and MCP servers/clients Risk, intent, and context evaluation for detections and response actions
Response & Recovery	Autonomous containment AI-assisted playbooks Forensics data capture for AI-related events Model rollback mechanisms Backup and restore for models and datasets	Kill switch for agents Autonomous response to agents performing risky behaviors Model and dataset rollback Remediation plans Tabletop exercises Supplier coordination plans Post-incident AI performance validation

AI security requires continuous visibility and behavioral detection

AI changes how fast systems move, how decisions are made, and how risk propagates. It does not change the fundamentals of security. Organizations that succeed will be the ones that apply those fundamentals rigorously, assume failure, and build systems that can detect, contain, and recover when AI behaves in ways they did not anticipate. Security is not what AI is allowed to do. It is whether the organization can understand, trust, and control what AI actually does in practice.

Take this guidance to understand different initiatives that organizations should be considering. Securing AI is the most critical component to AI safety. As organizations invest more in AI adoption, they should be investing in security in order to mitigate the risk of AI adoption. Organizations should be evaluating their governance and security stack to include well-integrated tools that are deployed, tested, operationalized and embedded within security workflows. While organizations mature in governance, visibility and identity access management, they should be investing in behavior-based detection and autonomous containment to mitigate AI risk.

About the author

Blog

July 6, 2026

NIST Just Proved It: AI Security Can’t Be Solved With Rules

Static AI guardrails are inherently limited

As organizations adopt generative AI, many still assume that the right set of guardrails will be enough. The problem is you can’t anticipate every way these systems might be misused, abused or attacked. What NIST has done is put a mathematical foundation under that intuition.

In recent research building on Gödel’s incompleteness theorems, which showed that any system built on a fixed set of rules will always have gaps, NIST demonstrates that there is no finite set of guardrails that can be universally robust against adversarial prompts. In plain terms, if your defense is based on a fixed set of rules, there will always be inputs that bypass them. Not because the rules are badly written, but because the problem space is bigger than static rules can ever cover.

This is not new in cybersecurity - detection rules have always had to live with this trade-off. What is different with GenAI is the scale and shape of that problem. These systems are built on human language, and human language is not bounded. It is fluid, contextual and deliberately ambiguous. The number of ways intent can be hidden is effectively limitless. You are not defending against a defined protocol or a fixed exploit chain. You are defending against the entire expressive capacity of people.

So attempting to create a complete set of rules is the wrong starting point. It assumes the problem can be deterministically described. NIST’s work shows that it cannot. Organizations still need a way to manage AI risk, but the traditional approach of defining allowed and disallowed patterns is always going to lag behind what is actually happening. The same input can be benign in one context and risky in another, and static rules struggle to capture that distinction.

The question then is what fills that gap?

AI security must shift from rules to behavior

What's required is a shift in what you are trying to understand. Rules try to describe what should and shouldn't happen. Behavior shows you what is happening. Or to put it another way, if inputs are unbounded and adversaries adapt, the only stable signal is behavior.

In a GenAI context, that means analyzing how an AI model is being used, how prompts evolve over time, how outputs are shaped, and where AI agent interactions start to drift from what is expected. It means moving from static definitions of bad to a more dynamic understanding of intent.

Instead of trying to predict every bad prompt, you focus on identifying when behavior starts to move outside expected norms. Instead of asking whether a single input matches a rule, you ask whether the overall pattern of activity makes sense for the system and how it’s being used.

Guardrails remain important but they are only one layer

This does not eliminate the need for guardrails. They still play a role. But they will never address the entire problem space and are simply one part of your defense in depth approach.

NIST’s proof is useful because it makes this explicit. It removes the assumption that with enough effort, a complete rule set is achievable. It isn’t.

Once you accept that, the shift becomes unavoidable. This is no longer a problem of writing better rules, but of understanding behavior in a space where the possible inputs are effectively unbounded.

For security leaders, that changes the nature of the problem. It is less about defining what should be allowed, and more about recognizing when something is no longer consistent with expected behavior.

That does not remove the need for guardrails, but it does change their role. They set boundaries, but they do not define understanding. The gap between the two is where risk now sits.

In the end, this is what “can’t be solved with rules” really means. Rules will always leave gaps, and those gaps are not theoretical. They show up in how systems actually behave Not what we expect them to do, or what we intended them to do, but what they are doing in practice. That is where the signal is, and increasingly, that is where the security problem sits.

References:

https://www.nist.gov/news-events/news/2026/06/nist-mathematical-proof-supports-transition-continuous-monitor-and-update

https://ieeexplore.ieee.org/document/11475847

About the author

Andrew Hollister

Principal Solutions Engineer, Cyber Technician

Your data. Our AI.

Elevate your network security with Darktrace AI

Get a demo

Check out this article by Darktrace: Post-Exploitation Activities on PAN-OS Devices: A Network-Based Analysis

Post-Exploitation Activities on PAN-OS Devices: A Network-Based Analysis

Introduction

CVE-2024-3400

Campaign Analysis

1. Exploit Validation and Further Vulnerability Enumeration

2. Command and Payload Transmission

a) Shell Script Retrieval

b) Executable File Retrieval

3. Configuration Data Exfiltration and Unusual HTTP POST Activity

4. Ongoing C2 and Miscellaneous Activity

Conclusion

Appendices ‍

Pre-CVE-Release IoCs

Indicators of Compromise

Darktrace Model Alert Coverage

MITRE ATTACK Mapping

References

From Click to Command: Behavioral Detection of AppleScript-Led MacOS Intrusions

Securing AI: Analysis of the Complete Security Stack with Governance and Controls

Enjoying the blog?

More in this series

From Click to Command: Behavioral Detection of AppleScript-Led MacOS Intrusions

Hola VPN Abuse: From Proxy Traffic to Malware and Cryptomining

How Attackers Abuse the Chinese Nezha Monitoring Tool

Blog

July 8, 2026

Securing AI: Analysis of the Complete Security Stack with Governance and Controls

Why traditional cybersecurity approaches are not enough for AI

How NIST defines the three core pillars of AI security

Secure

Defend

Thwart

How this looks in practice

Why agentic AI fundamentally changes enterprise cyber risk

Behavioral risk

Structural risk

Accountability

What does defense in depth look like for AI?

How security leaders should prepare for enterprise AI adoption

What capabilities should every AI security program include?

Layered Governance and Security Stack for Securing AI

AI security requires continuous visibility and behavioral detection

Blog

July 6, 2026

NIST Just Proved It: AI Security Can’t Be Solved With Rules

Static AI guardrails are inherently limited

AI security must shift from rules to behavior

Guardrails remain important but they are only one layer