New findings covering auto-contextualized mobile logic bombs, blurred capability boundaries in LLM applications, side-channel attacks on serverless cloud functions, cross-compilation binary alignment, and function name inference in stripped binaries.
New findings spanning library fuzzing, BACnet protocol vulnerabilities, OLE document security, backdoors in machine learning, risks in AI character platforms, and cloud app-in-app resource isolation.
Misconfigured email auto-configuration globally, manipulating GitHub contribution histories to infiltrate supply chains, and adversarial Android malware factories.
Misaligned models evading blocking monitors via prompt injections, a semantic cover story attack that reframes malicious behavior, inference engine fingerprinting by models, and a look at temporal graph reachability vs. observation.
New findings on the attack surface of LLM agents, from cross-channel fragmentation in tool-calling pipelines to adversarial task contamination and supply chain poisoning of self-modifying code.