Backdoor Attacks on Decentralised Post-Training

This collection of papers (arXiv 2604.*) highlights multiple emerging security risks and some corresponding mitigations in AI, ML, and cyber systems. Key threats: a novel backdoor attack on pipeline-parallel post-training where an adversary controlling an intermediate pipeline stage can inject alignment-breaking backdoors; environment-injected trajectory-based memory poisoning (eTAMP) that contaminates web-agent memories via passive observation to achieve cross-session, cross-site compromise; safety “unalignment” techniques (weight orthogonalization and jailbreak-tuning) that meaningfully re‑e