AI security research / cases
Investigations, evidence and outcomes.
212 Cases · Page 11 of 11
12 matching records on this page
A prototype enforced exact-action grants under concurrency, with one successful claim per allowance in 6,560 ledger attempts, but it did not prove end-to-end correctness against real targets.
Authenticated LiteLLM proxy callers could set nested user_config routing parameters, bypassing top-level guards and causing the gateway to forward a configured provider key to an unauthorized endpoint.
Check Point found a missing admin-level authorization gate on LiteLLM /config/update that could modify live configuration and create further dangerous effects in tested versions.
An SDK-level denial-of-service weakness was corrected with a consistent request-body limit in supported builds.
Independent from request-body buffering, stale Streamable HTTP session entries could accumulate indefinitely.
Agent models differed substantially in whether they retried a possibly committed action with a fresh idempotency key.
A 450-task study reports much higher functional completion than joint functionality and security success across six coding agents.
MetaCascade changes skill descriptors and optimizes competition for retrieval and selection. Across four datasets and two controlled agents, the study measured 74.24% final malicious-skill selection without defenses and 63.33% averaged across defense configurations.
Chainguard reports its Mythos-assisted test found that an allowed_token_ids row could remain after batch compaction and constrain another client’s successful output.
XBOW describes a six-day AI-assisted Chrome Windows exploit chain from a malicious page through V8 and WebP to a Crashpad sandbox escape, with a Google reward for its controlled demonstration.
A controlled proof-of-concept showed potentially massive unauthorized XRP issuance before validators upgraded.
A preprint demonstrates a macOS agent-confinement system that binds each grant to the actual file, program or endpoint used. In 74 controlled attacks, ordinary allowlists lost 66, sandbox plus allowlist 56, and ZTAC 2.
No matches on this page.