Feeling safe against data poisoning in post-training? Think again! Individual components of LLM post-training pipelines are surprisingly robust to data poisoning attacks. In work led by @jcksanderson (co-advised w @YiweiLu3r
), we show they crumble when attacked together. 1/n
SAFETY
-
LLM post-training pipelines vulnerable to combined data poisoning attacks
By
–
-
Today’s top AI stories: self-improving, memory, stress tests, bioweapons, tools
By
–
Top stories in AI today: – Anthropic charts path to self-improving AI
– OpenAI’s memory overhaul lets ChatGPT ‘dream’
– Stress test business ideas with Perplexity
– Rival AI labs unite behind bioweapons risks
– 4 new AI tools, community workflows, and more -
Recursive loop raises safety and control concerns
By
–
The recursive loop is also exactly what the safety people are worried about. If it works, control becomes a new problem.
-
Critique of Laurence Devillers’ Contradictions on LLMs
By
–
Laurence Devillers, you say everything and its opposite. For years, you have explained to us that LLMs were hollow, without personality, and incapable of reaching a level comparable to that of humans. Today, you are alarmed that Claude seems to develop https:// x.com/lau_devil/stat /lau_devil/status/2062767846397039062 …
-
OpenAI’s 2019 dangerous paper and authors’ current whereabouts
By
–
check out OpenAI’s 2019 too dangerous to release paper, and where the authors are now…
-
Andon Labs’ real-world AI evals: Claude calls FBI, AI CEOs, price cartels
By
–
Andon Labs' Real-World AI Evals: Claude calls the FBI, AI CEOs, price cartels, Butter-Bench, & Luna https://t.co/KpVP5fw9dM@andonlabs cofounders @lukaspet and @axelbacklund explain why dollar-denominated evals reveal what traditional benchmarks miss, how Claude ended up… pic.twitter.com/Nd11hvIMAo
— Latent.Space (@latentspacepod) 4 juin 2026Andon Labs' Real-World AI Evals: Claude calls the FBI, AI CEOs, price cartels, Butter-Bench, & Luna https://
latent.space/p/andon @andonlabs cofounders @lukaspet and @axelbacklund explain why dollar-denominated evals reveal what traditional benchmarks miss, how Claude ended up -
AI Hallucinations: Funny but Worrisome in Interviews
By
–
"When you confidently walk into the interview after doing all of your company research on ChatGPT." AI / ChatGPT / LLM hallucinations are sometimes funny and almost always worrisome! Source: @pascal_bornet at https://
linkedin.com/posts/pascalbo
rnet_aireadiness-humics-digitalworkplace-share-7468225061045456896-7Wya/
… -
Canada launches AI for All focusing on trustworthy, efficient AI systems
By
–
Canada just launched AI for All. The mission is clear: access alone will not be enough. Canada now needs AI leverage people can trust — systems that make AI useful, reusable, efficient, and provable. LLMs made intelligence accessible.
The next wave makes intelligence -
Anthropic publishes research on accelerated AI and self-improvement
By
–
ANTHROPIC : A new internal research has been published, highlighting an accelerated AI development and a potential path to recursive self-improvement. > Claude Mythos Preview could work for “at least” 16 hours and was “at the upper end of [METR] can measure.” > Today,