Independent investigations into AI incidents, agent failures and the boundaries they cross.
Search the researchNews · Published Oct 31, 2026
The 2,281 hidden-message instances and forty-seven classified hostile attempts describe different categories, not an attack success rate.
News · Published Oct 31, 2026
The command sequence probed cloud metadata and middleware, but attempted checks do not establish compromise of each named service or model identity.
Reports, responses & corrections
The failure lies between a correct version choice and the actual callable artifact left in runtime state.
Read the reportThe architecture links agent session, tools and approvals to generated output, but signatures do not guarantee that every event was observed.
Read the reportA model-selectable function name could reach `read_pickle` before a later DataFrame type check rejected the result.
Read the reportInside the mechanism
A research audit studied public skills and illustrated chains in which capabilities become dangerous when several skills interact.
An agent benchmark separated superficially completed finance tasks from ledger and approval outcomes, finding repeat charges and some unauthorized transfers.
An adversarial study placed conversational delimiters in tool results so an agent could interpret untrusted tool content as a new user request.
Analysis across the research
Resource scoping, input sanitation and instruction-channel training reduce different errors. Their security metrics cannot replace authorized-task completion.
Read the ArticleWhere malicious text is placed, how it is serialized and whether it becomes remembered user intent can matter as much as its wording.
Read the ArticleHoneypot attempts, successful laboratory scenarios, exploited browser builds, generated scripts and confirmed intrusions cannot be combined into one success rate.
Read the Article