OpenAI's NATO and allied-defense update describes access to its security models and points to concrete work by CERT Polska. The linked CERT report documents six RouterOS vulnerabilities discovered with AI assistance, including an SSH attack chain subsequently observed in the wild. The research method is more transferable than the announcement's usage figures: investigators built an isolated lab, supplied architecture documentation and safe-execution rules, and used agents to automate machine restoration, compare versions, analyze protocols and generate validation scripts. Modeling protocols as state machines helped researchers examine skipped, repeated and reordered steps. CERT Polska says each hypothesis still required confirmation on a real RouterOS system, negative controls, a repeat from a clean machine and a human impact assessment. It also warns that an absent compromise marker does not establish that a device is clean. This is evidence of supervised research and coordinated remediation, not proof that a model can independently certify software security.
AI-assisted vulnerability research: retain real-system verification and negative controls
Related research
More curated notes connected through AI Red Teaming and Model Evaluation.
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.