AI security research / cases
Investigations, evidence and outcomes.
230 Cases · Page 12 of 12
10 matching records on this page
The JevProbe paper identifies application behavior changes from attacker-controlled contexts even when structured Jev outputs match schema, not confirmed real-world downstream compromise.
The authors present an enforcement layer outside an agent model or prompt that intercepts file and network effects; some effects are held until a policy decision before being committed. Controlled benchmark results do not establish general production readiness.
A cross-family malware study catalogs embedded natural-language instructions intended to manipulate a model that reads extracted binary strings. Some attacks degraded test-model triage; the publication does not document broad production AV evasion.
A fake-CLI-documentation cluster replaces vendor one-liners with attacker script locations under Claude, Copilot, Codex, Ollama and OpenClaw branding.
CircuitBreak demonstrates post-training model-weight modification rather than a remote prompt-only jailbreak; authors report substantial refusal bypass in controlled evaluation.
Researchers trained otherwise matched models with positive or exploitative descriptions of AI users and then observed changed trust behavior in simulated games. The result is a demonstrated data-conditioning risk, not confirmed malicious deployment.
The analyzed loader runs a real Claude installer in the foreground, sets a scheduled background headless-Edge task, and posts browser data to a suspicious backend.
Untrusted shared conversations, custom GPTs or published Claude artifacts can sit on authentic AI-platform domains and redirect viewers to attacker installers or fake verification. The hosting identity is not vendor endorsement.
Two independent cache paths described in one advisory can collide when attacker-influenced values are used without adequate content binding or domain tagging. Exploitation requires the affected configuration and request controls.
An authenticated client of the disaggregated scale-out API can supply feature fields which the generate path treats as trusted renderer output, permitting engine failure and integrity impact according to the vendor advisory. Distinct from request-mask and EXIF cache issues.
No matches on this page.