AI security research / cases
Investigations, evidence and outcomes.
288 Cases · Page 15 of 15
8 matching records on this page
A repository role-authorization fallthrough let some Viewer or unassigned accounts perform write operations intended for Editor or Admin roles.
A reachable terminal WebSocket could replay a named PTY output buffer before verifying its client token, making output secrecy different from command-input blocking.
Isolated testing found 150 filesystem-boundary violations across a selected sample of MCP servers, including 136 disclosures of synthetic planted credential markers.
Researchers show shared GPU page-migration counters and driver scheduling can cross GH200 MIG tenant boundaries despite isolated GPU memory.
Hunt.io analyzed an active PaperCut attacker toolkit with agent instructions, distinct exploit paths and confirmed post-compromise reverse SOCKS infrastructure; raw scanning totals remain separate.
Adaptive 4B auditors probed a 30B flight-booking agent for multi-turn policy failures, yet a human audit found 121 false positives in 207 model-judged policy instances.
A first-hand support report records Perplexity decisions on obvious known-answer prompts nearly collapsing to uniform options while HTTP requests still succeeded; provider acknowledged an issue and a later recovery.
The released SPACE agent sandbox platform separates Firecracker microVM isolation from distributed routing and per-execution authorization to reduce cross-sandbox confusion.
No matches on this page.