AI Security Incident Roundup: November 2026

Covers November 1–22, 2026. Details reflect public reporting available through November 22 and may change as investigations continue.

The common thread this month is that the weak point was rarely the model itself. It was the infrastructure around it: inference servers, vendor keys, extension marketplaces, application builders, software dependencies and the logs where secrets quietly accumulate. AI also appeared on the attacker side in a more concrete way, particularly in automated vulnerability discovery and exploitation.

For defenders, these developments reinforce several important priorities: securing AI infrastructure, controlling third-party dependencies, protecting credentials, monitoring agent activity and reducing the time between vulnerability disclosure and remediation.

Each item below includes a summary, practical defensive implications and links to further reporting. The analysis is intended to help security professionals understand the implications rather than simply repeat headlines.

This edition covers November only. For the October PHP2Shell disclosure, see our earlier article, PHP2Shell: How AI Helped Uncover Critical Vulnerabilities.

At a glance

Item When One-line lesson
NEAR AI breachNov 8A secure enclave does not protect against bugs in the software running inside it.
Fireworks AI to Cursor incidentOct 29, disclosed Nov 13Agent traces and vendor keys are credential stores.
Grok Build appsNovemberLink-only sharing is not access control.
Agentic attack frameworksNov 13–17Automated exploitation can operate at machine speed.
Stolen AI keysNovemberSpend limits and scoped credentials are essential safeguards.
Lazarus researchNov 10Behavioral tests cannot prove model weights are clean.
Microsoft Agent Framework compromiseNov 17AI software dependencies can become credential-stealing entry points.
LiteLLM vulnerabilityNov 18AI gateways require the same security attention as critical infrastructure.
Cairn and MONAI LabelNov 20Autonomous agents can discover and exploit application vulnerabilities.
European Commission cyber incidentNov 19AI-assisted cyber campaigns are a concern for public institutions too.

1. NEAR AI: The enclave did not cover the software inside it

What happened. On November 8, an attacker exploited a remote code execution flaw in SGLang, the open-source inference engine, to run code on one NEAR AI inference server. From there they reached an internal system behind a retired Private Chat product and downloaded account details for about 100 users. They also reached 18 agents on a legacy deployment that ran outside the company's confidential computing environment.

NEAR AI says requests being processed on the compromised server that day, including ones in confidential mode, could have been observed. It found no sign that stored chats or uploaded files were accessed. Two customers had API keys sitting in logs, and those keys have been revoked. Cloud Security Alliance's summary notes that NEAR AI has not said which SGLang flaw was used.

In the same week, separate research on AMD's SEV-SNP showed that a malicious host can read protected guest memory, with full fixes depending on vendor BIOS updates expected in November and December.

What this means for defenders.

  • Treat the inference engine as part of your trusted computing base. Confidential computing protects data from the host and operator. It does not protect against vulnerabilities in code running inside the protected environment.
  • Ask vendors the unglamorous questions. How quickly do they patch SGLang or vLLM? What is exposed to the network? What egress is allowed? How are secrets kept out of logs?
  • Understand attestation limitations. Attestation shows which build is running, not that the build is free of vulnerabilities.
  • Retire old services properly. Anything left running after a product is shut down remains potential attack surface.
  • Keep secrets out of logs. Two customers' keys reportedly leaked this way.

Sources: NEAR AI incident report · Cloud Security Alliance briefing, Nov 11 · Cloud Security Alliance briefing, Nov 16

2. Fireworks AI to Cursor: A vendor key becomes an IDE backdoor

What happened. On November 13, Cursor reported that its October 29 marketplace incident began with an access key it had issued to AI infrastructure provider Fireworks AI, which was exposed in a Fireworks breach. According to the reporting, the attacker used the key to copy historical coding-agent traces that still contained user credentials, used those to reach Cursor's internal tooling, and then placed two tampered copies of Anthropic's Claude Code extension (versions 2.1.312 and 2.1.313) in the marketplace.

They were live for roughly 11 hours, reached at most about 5,000 users, and tried to plant a persistent SSH backdoor. Anthropic's own releases and other marketplaces were reported as unaffected. Fireworks separately widened its disclosure on November 9 to cover customer datasets, secrets and partial model weights.

What this means for defenders.

  • Agent traces are credential stores. Redact secrets before traces are logged or shared with a vendor, and set short retention limits.
  • Inventory vendor keys. Every integration needs an owner, access restrictions and a rotation plan.
  • Control IDE extensions. Pin versions, restrict auto-update and prefer trusted distribution sources.
  • Investigate persistence. If a machine loaded affected versions, check for unexpected SSH keys and other indicators. Uninstalling alone may not remove persistence.

Sources: Cursor incident update · BleepingComputer · The Hacker News · Fireworks AI bulletin

3. Grok Build apps: Public by default

What happened. Security firm Claroc reported that 35,578 of 37,571 live apps built with Grok Build, approximately 95 percent, could be opened without signing in. Roughly one in ten reportedly held internal data, including a hospital staff roster of more than 3,500 people and plaintext passwords.

Apps defaulted to sharing with anyone who has the link, and their subdomains were visible in passive DNS records, making supposedly hidden links discoverable.

What this means for defenders.

  • AI app builders are publishing channels. Employees can create and distribute internal tools outside traditional IT review.
  • A hard-to-guess link is not access control. Require authentication and review publishing permissions.
  • Monitor exposed data. Monitor certificate transparency and passive DNS for unexpected application domains associated with your organisation.
  • Set data classification rules. Tell employees what information may be used in AI-built applications.

Sources: Claroc research · The Hacker News · Cloud Security Alliance briefing

4. Agentic attack frameworks go operational

What happened. On November 13, the FBI warned that an open-source agentic penetration-testing framework called ARTEX had been used against organisations in finance, government, technology, defence, healthcare and energy since at least August.

Amazon, Cloudflare and Palo Alto Networks' Unit 42 each reported intrusions involving ARTEX or similar frameworks, and those agents tend to go after AI gateway credentials first.

Then on November 17, Google's Threat Intelligence Group and Mandiant reported what they describe as the first case they have seen of an LLM framework finding and exploiting vulnerabilities with no human in the loop. An exposed staging server ran a multi-agent setup that compared the code of popular web applications before and after patch announcements, worked out the flaw, drafted an exploit and ran it.

Google's September report had said it had not yet seen fully autonomous pipelines in the wild, making this a notable development. Google also reported that vulnerability disclosures and exploits nearly doubled between January and August 2026.

What this means for defenders.

  • Patch-diffing is being automated. Assume published patches can provide attackers with a map to the underlying vulnerability.
  • Detect behaviour, not tool names. Frameworks evolve quickly. Look for machine-speed request patterns, unusual reconnaissance and callbacks to LLM APIs.
  • Harden AI gateway credentials. Default or weak keys on gateways such as LiteLLM can become early targets.
  • Shorten patching windows. Internet-facing systems need rapid remediation processes for critical vulnerabilities.

If you log AI application traffic, our guide to detecting prompt injection in application logs covers the fields worth capturing.

Sources: FBI IC3 alert · Help Net Security on Google's report · Help Net Security on ARTEX · Amazon Threat Intelligence · Cloudflare · Unit 42

5. Stolen AI keys: Who pays the bill?

What happened. Railway estimated that keys stolen from its customers drove about $740,000 in fraudulent AI usage. One leaked Gemini key reportedly ran up more than $82,000 in roughly 48 hours.

Google's Mandiant team described several 2026 cases where fraudulent charges passed $100,000 in a single day, with the worst reaching about $1 million before it was stopped. The stolen access is resold through underground storefronts that sell frontier and jailbroken models by the token.

What this means for defenders.

  • Set hard spend caps. Configure provider-side usage limits and budget alarms rather than relying only on after-the-fact billing alerts.
  • Scope credentials. Restrict each key to specific models and applications, rotate regularly and test revocation procedures.
  • Review contracts and insurance. Understand who carries financial responsibility for unauthorised usage.
  • Check where keys leak. Logs, traces, repositories and development environments all need secret-scanning controls.

Sources: The Hacker News on Railway · The Register · Help Net Security on LLMjacking · Cloud Security Alliance briefing

6. Research watch: Lazarus, a backdoor made from a handful of neurons

What happened. On November 10, Microsoft Research described a technique called Lazarus. By rewriting fewer than ten rarely used neurons in an open-weight model, without retraining, a person with write access to the model file can plant a trigger that makes an agent emit an attacker's command.

In their main demonstration they changed seven of more than 12,000 neurons, and the modified model still behaved normally on standard tests. Microsoft reported no use in real attacks. Its detection methods flagged about 91 percent of held-out modified checkpoints, with a false-positive rate near 10 percent.

What this means for defenders.

  • Behavioural testing cannot prove a model is clean. A backdoored checkpoint can pass evaluations.
  • Verify provenance. Download from trusted publishers, pin hashes and compare checkpoints to original publisher versions where possible.
  • Limit agent permissions. Least privilege, sandboxed code execution, restricted outbound connections and human approval for high-impact actions reduce potential damage.

Sources: Microsoft Research · Microsoft Security · CSO Online · InfoWorld

7. Microsoft Agent Framework: Malicious PyPI releases steal developer credentials

What happened. On November 17, two malicious versions, 1.26.1 and 1.26.2, of Microsoft's official agent-framework Python package were published to PyPI. The releases contained credential-stealing code that ran through Python startup hooks. This meant the malicious code could execute whenever Python started in an affected environment, even if the framework was never imported.

The affected versions were subsequently quarantined, with version 1.26.0 identified in the reporting as the last known clean release.

What this means for defenders.

  • Check dependencies. Search lockfiles, build logs, virtual environments and container images for versions 1.26.1 and 1.26.2.
  • Investigate affected environments. Rebuild them from trusted sources and review potential credential exposure.
  • Rotate credentials. Consider cloud keys, developer tokens and CI/CD secrets accessible to affected environments.
  • Secure package publishing. Apply strong publisher authentication, protected release workflows and independent verification of package artifacts.

Sources: Wiz investigation · BleepingComputer · Cloud Security Alliance research note

8. LiteLLM: Vulnerability added to CISA KEV

What happened. On November 18, CVE-2026-40217 affecting BerriAI LiteLLM was reported as added to the CISA Known Exploited Vulnerabilities catalogue. The vulnerability involves an unprotected alternate channel that can allow remote code execution.

This is particularly relevant because LiteLLM is used as an API gateway for connecting applications to different LLM providers. A compromise of this layer can affect multiple AI services and the credentials used to access them.

What this means for defenders.

  • Identify exposed deployments. Inventory LiteLLM instances, particularly internet-facing gateways.
  • Apply vendor remediation. Review the official security advisory and implement the recommended fixes.
  • Restrict administrative functionality. Limit access to management and custom-code endpoints.
  • Review activity logs. Investigate suspicious requests and unexpected code execution.
  • Protect AI gateways. Treat them as critical infrastructure rather than ordinary application middleware.

Sources: CISA Known Exploited Vulnerabilities catalogue · LiteLLM security advisories

9. Autonomous AI agents discover and exploit a vulnerability in medical imaging software

What happened. On November 20, Amazon Threat Intelligence published findings about a Chinese-nexus threat actor using the Cairn autonomous penetration-testing framework. The agents were given a broad objective and independently performed reconnaissance, vulnerability testing and exploitation.

In one operation, the agents reportedly discovered a path traversal vulnerability in MONAI Label, an open-source medical imaging AI application. They used it to obtain file-write access and subsequently establish a root-level reverse shell. The vulnerability was assigned CVE-2026-133968. Amazon reported that the project maintainers were notified and released a fix.

What this means for defenders.

  • Prepare for autonomous offensive AI. AI-driven exploitation is an emerging operational threat, not just a research possibility.
  • Prioritise internet-facing systems. Patch exposed applications and development environments quickly.
  • Validate file handling. Review upload endpoints, path traversal protections and file access controls.
  • Monitor attack behaviour. Look for unusual reconnaissance, API requests and file operations.
  • Apply least privilege. Isolate AI-enabled workloads and restrict access to sensitive resources.

Sources: Amazon Threat Intelligence · NIST National Vulnerability Database

10. European Commission confirms an AI-assisted cyberattack

What happened. On November 19, the European Commission publicly confirmed that it had been targeted by an AI-assisted cyber campaign detected on October 22. Several applications were temporarily taken offline as a precaution, with most subsequently restored.

The Commission said its investigation had not found evidence of a data breach affecting staff as a whole. The attacker and full scope of the incident had not been publicly established at the time of reporting.

What this means for defenders.

  • Include AI-assisted threats in incident planning. Attackers may use AI to accelerate reconnaissance and other stages of an intrusion.
  • Protect critical services. Maintain application availability, recovery procedures and tested response plans.
  • Improve monitoring. Ensure suspicious activity across applications and infrastructure can be correlated.
  • Be careful with attribution. Avoid assigning responsibility without supporting evidence.

Sources: Reuters · European Commission audiovisual service · Associated Press

Policy signals

Three regulatory and policy developments are worth monitoring, although their practical implications and implementation timelines may evolve.

  • Australia: Australia accepted all 18 recommendations from its review of an OpenAI agent's unauthorised access to a government statistics portal. The plans include a new AI safety regulator and time-bound reporting of AI incidents, including misaligned agent behaviour. Legislation is expected in 2027.
  • ENISA: ENISA's November 10 state-of-cybersecurity report argues that as AI speeds up vulnerability discovery, triage and remediation capacity is becoming a key bottleneck.
  • US Federal Reserve: Its November 13 Financial Stability Report found that 87 percent of surveyed market contacts now cite AI as a salient risk, up from 50 percent in the spring.

For defenders: Map agent telemetry to potential incident-reporting categories, check vendor contract notification requirements and use these developments to frame AI security investments in terms of operational and business risk.

Sources: The Next Web on Australia · ENISA · Federal Reserve Financial Stability Report · CNBC

Checklist for December

These incidents highlight practical controls organisations should review before the end of the year.

  • Redact secrets from logs and agent traces, and limit how long vendors can retain them.
  • Inventory every key issued to vendors and every AI provider key in use. Scope, cap and rotate credentials.
  • Pin and review IDE extensions, Python packages and model downloads like any other software dependency.
  • Treat inference engines and AI gateways as part of the trusted computing base and patch them quickly.
  • Shorten patch targets for internet-facing systems and maintain an emergency remediation process for critical vulnerabilities.
  • Decommission retired AI products and legacy agent deployments fully.
  • Require authentication by default on applications published through AI app builders.
  • Verify model provenance and integrity before deploying externally sourced checkpoints.
  • Restrict agent permissions, network access and high-impact actions.
  • Review incident response plans for AI-assisted reconnaissance and automated exploitation.

For a fuller control list, see our AI Security Checklist, and for how agent risks fit together, see the OWASP Top 10 for Agentic Applications overview.


Final perspective

November's developments demonstrate that securing AI systems requires more than testing model behaviour. The surrounding infrastructure, software dependencies, access credentials, deployment environments and agent permissions all form part of the security boundary.

For security architects and engineering teams, the priority should be to apply established security principles to these new systems: secure-by-design architecture, least privilege, dependency integrity, strong identity controls, continuous monitoring and effective incident response.

This roundup summarises public reporting and adds independent analysis for educational purposes. It is not a substitute for a security assessment of your own environment. Readers should consult the linked primary sources for the latest technical details and remediation guidance.

Comments