Evaluating Generalization Mechanisms in Autonomous Cyber Attack Agents

Collection of March 12, 2026 security-relevant arXiv papers covering attacks, defenses, and supply-chain/agent governance for modern AI and distributed systems. Key findings: Flip-Agent demonstrates the first targeted bit-flip attack (BFA) framework against multi-stage LLM-based agents, enabling manipulation of outputs and tool invocations and exposing a critical hardware/fault-injection attack surface for agent pipelines. NabaOS proposes a practical, low-latency HMAC-signed tool-receipt scheme to detect agent hallucinations in real time as an alternative to expensive zk-proof approaches. SBOM