Google Threat Intelligence Group and Mandiant report adversaries using multi-agent frameworks during live intrusions and automating parts of vulnerability research. A key example is an exposed staging server containing a framework that compares software versions, reasons about patches, prepares exploits and executes them; the server also contained victim data and post-compromise tooling. The report distinguishes observed deployments from experimentation with other frameworks and warns that many published AI-generated exploit claims are nonfunctional. It also describes a suspected North Korean cluster modifying repository-local Claude Code settings so that starting the agent triggers backdoor delivery after an initial compromise. These findings connect familiar weaknesses—exposed credentials, unpatched services and trusted executable configuration—to faster attack workflows. Google notes that such tooling can leave conspicuous traffic and code artifacts, and that model errors can undermine attacker reliability. The evidence supports specific observed automation, not a claim that every intrusion is fully autonomous or stealthy.
GTIG: shorten patch exposure and inspect agent configuration as executable code
Related research
More curated notes connected through Agent Security and AI Red Teaming.
Pacing model development in an era of cyber-critical capabilities
OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.
OWASP Top 10 for Agentic Applications for 2026
OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.
A blueprint for democratic governance of frontier AI
OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.