Tanmay and Dolly Sah examine a failure ordinary coding benchmarks miss: an operation commits, its acknowledgment is lost, and an agent retries it. UndoBench pairs normal and faulted runs with matching seeds and checks both final state and external effect history. Its frozen study covers 12 held-out workflows and 2,880 paired trials; normal success was 83.54%, while recovery among normally successful pairs was 46.72%. Those percentages have different denominators. The experiments use sandboxed services and one injected fault per trajectory, limiting claims about production reliability. The companion EvoUndo research captures pre-change state, declares affected state and independently checks recovery before admitting a persistent harness modification. Its scope is modeled harness state; third-party APIs and distributed external state are not modeled. Effect-scoped snapshots performed better in the paper's serializable-state comparison. Together, the work supports testing recovery at specific mutation boundaries rather than assuming that a successful task, a Git revert or another prompt can safely reverse every effect.
UndoBench and EvoUndo: test recovery after side effects, not just task completion
Related research
More curated notes connected through Agent Security 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.