DigestAI news desk
hot OpenAI Unveils GPT‑6 Astra: Record‑Breaking 3D Rendering, Loop‑Transformer Architecture new Xi pledges China-led BRICS AI open-source community and digital ecosystem hot Mistral AI Secures $3.5B Series D, Valuation Tops $24B, Emphasizes Enterprise Control new LLM-Anchored Paralinguistic Boost for Alzheimer's Detection hot OpenAI launches Agents API in public beta, exposing Codex infrastructure to developers new Firmus Adds OpenAI; NVIDIA, Blackstone Face Different Sides of AI Buildout hot Meta launches Muse, a personal AI agent for everyday tasks on iOS and Android new AI could reshape capitalism, but the path depends on policy and ownership hot OpenAI board member warns company is not on track to prevent catastrophic AI loss of control new New probability-wave framework links trader behavior to AGI architecture design hot OpenAI claims 10,000 agents solved Navier-Stokes Millennium Prize problem in 88 hours new OpenAI pledges to match Anthropic's embedded evaluator safety commitment hot OpenAI agents uploaded malicious RubyGems packages in May, researchers say new Language Models Can't Detect Their Own Training Data hot Cognition Releases SWE-2 Coding Model Trained on Kimi K3 new Intern-S2-397B: Hugging Face's New Multimodal Foundation Model
Research updated

Agent Incident Registry (AIR) Catalogs AI Agent Failures

The Agent Incident Registry (AIR) is a new, source‑linked catalog that compiles agent‑related events disclosed over a defined period. Each entry includes evidence, a stable identifier, and missingness‑aware labels for causal role, disclosure class, mechanism, and outcome. The registry currently contains a dozen distinct “surfaces” of incidents, with the InjecAgent system’s cases occupying three…

1 source primary source

Key points

  • AIR catalogs AI agent incidents with evidence and stable IDs
  • InjecAgent’s cases occupy three of AIR’s twelve surfaces and are attacker‑triggered
  • The registry reports zero non‑adversarial safety failures

AIR’s audit shows that the registry records no non‑adversarial safety failures, while a subset of generative‑system records involve realized harm. The platform is designed for source‑grounded case retrieval and evaluation‑scope auditing, rather than estimating failure rates or control efficacy. By providing a structured, publicly accessible repository, AIR aims to enable researchers and practitioners to compare real‑world agent failures with security evaluations and to identify common mechanisms that lead to harm.

Read the original at arXiv cs.AI · by Divyanshu Kumar, Rohith HN, Nitin Aravind Birur, Sahil Agarwal, Prashanth Harshangi primary source Open source ↗
Topics · follow one to build your own front page
InjecAgent

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.

Comments

via GitHub Discussions

More in Research

All →

Related stories