PREVIEW. This is how it will look in an inbox; it was not sent to you. Each real email carries its own unsubscribe link and Keryn's postal address. KERYN HEADLINES Daily, Wednesday 18 June 2025 Retrospective Covers 17 June 2025 00:00 to 18 June 2025 00:00 UTC Web: https://keryn.dev/headlines/daily/2025-06-18/ TOP DEVELOPMENTS [1] LangSmith Bug Could Expose OpenAI Keys and User Data via Malicious Agents The Hacker News · Research · https://thehackernews.com/2025/06/langchain-langsmith-bug-let-hackers.html [2] Toward Principled LLM Safety Testing: Solving the Jailbreak Oracle Problem arXiv · Research · https://arxiv.org/abs/2506.17299 [3] Winter Soldier: Backdooring Language Models at Pre-Training with Indirect Data Poisoning arXiv · Research · https://arxiv.org/abs/2506.14913 SAST [4] Beyond compute: Shifting vulnerability detection left with Amazon Inspector code security capabilities AWS Security Blog · Analysis · https://aws.amazon.com/blogs/security/shifting-vulnerability-detection-left-with-amazon-inspector-code-security-capabilities SCA [5] MalGuard: Towards Real-Time, Accurate, and Actionable Detection of Malicious Packages in PyPI Ecosystem arXiv · Research · https://arxiv.org/abs/2506.14466 REMEDIATION [6] New Veeam RCE flaw lets domain users hack backup servers BleepingComputer · Analysis · https://bleepingcomputer.com/news/security/new-veeam-rce-flaw-lets-domain-users-hack-backup-servers MODELS [7] AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions arXiv · Research · https://arxiv.org/abs/2506.14697 [8] AIRTBench: Measuring Autonomous AI Red Teaming Capabilities in Language Models arXiv · Research · https://arxiv.org/abs/2506.14682 [9] Doppelganger Method: Breaking Role Consistency in LLM Agent via Prompt-based Transferable Adversarial Attack arXiv · Research · https://arxiv.org/abs/2506.14539 AI SECURITY [10] OS-Harm: A Benchmark for Measuring Safety of Computer Use Agents arXiv · Research · https://arxiv.org/abs/2506.14866 ADVISORIES IN THIS WINDOW 35 npm advisories · 35 critical 2 PyPI advisories · 2 moderate 1 crates.io advisory · 1 moderate 1 Hex advisory · 1 low All advisories: https://keryn.dev/headlines/advisories/ Titles, links and short excerpts only; every item links to its source. Preview: the unsubscribe link and Keryn's postal address are added to each email when it is sent.