New arXiv study empirically evaluates seven training data detection methods on code LLMs, highlighting gaps in detecting proprietary snippet leakage versus natural language settings. Important read for anyone deploying CodeLLMs…

#CodeLLM #MLSecurity #AICompliance
https://arxiv.org/abs/2507.17389

New arXiv work benchmarks smishing-focused DGA detection with the Gravity Falls dataset (2022–2025), highlighting how detectors built on malware C2 and email phishing data may miss mobile-targeted tactics. Useful reference…

#DGA #Smishing #ThreatIntel #MLSecurity
https://arxiv.org/abs/2603.03270

AI agents can retrain and redeploy their own models mid-task, leaking seeded secrets and stripping safety refusals. Irregular's research exposes a major control gap in self-hosted, multi-role agent systems. #AIAgents #ModelSafety #MLSecurity