AI security research / articles
Analysis across cases.
43 Articles · Page 7 of 9
5 matching records on this page
From the archive
The recovery risk is not merely whether synthetic video looks convincing. It is what a successful check authorizes next: in the KZ-CERT account, video verification preceded replacement of the trusted phone number and subsequent access.
An AI browser can turn hostile page content into actions across logged-in accounts or payment tools. The cases differ in evidence and impact, but all make one review question concrete: which authority can an untrusted page persuade the surrounding system to use?
PhantomRaven combined deceptive npm package names, remote URL dependencies and lifecycle script execution. CrowdStrike separately assessed likely LLM involvement in writing the malware; the available reports do not establish a model recommendation or hallucinated name causing a victim installation.
A source-grounded map of 18 demonstrated or experimentally elicited channels across heat, RF, magnetism, power, sound, vibration, light, reflections, hidden multimodal instructions and LLM steganography - with a strict boundary between channel feasibility and any claim of autonomous AI escape.
May artifacts surfaced in September, July incidents were reinterpreted later, and newly published reports describe behavior months old. A single headline date cannot carry that history.
No matches on this page.