12 stories
In the first of three posts on Opus 4.7, Mowshowitz walks through the 232-page model card's RSP evaluations, alignment risk assessment, and economic capability index, while flagging that model welfare and capabilities are held for separate posts.
In the third of three posts on Opus 4.7, Mowshowitz argues that Anthropic's model welfare evaluation process may have produced inauthentic self-reports by teaching the model to produce desired responses rather than accurate ones, drawing a parallel to how humans learn to mask internal states under institutional observation.
A METR staff member spent three weeks red-teaming a subset of Anthropic's internal agent monitoring and security systems, finding several novel vulnerabilities, some of which have since been patched.
METR researchers fine-tuned four open-source reasoning models on around 950 examples and found out-of-distribution chain-of-thought controllability rose from an average of 2.9% to 8.8%, suggesting the capability is latent rather than absent.
In Import AI 452, Jack Clark covers a Lyptus Research paper examining how AI performance on cyberoffense tasks has scaled since 2019, with capability doubling times tightening to 5.7 months among models released since 2024.
The Verge, citing Bloomberg, reports that a small group accessed Anthropic's withheld Mythos cybersecurity model through a combination of an educated guess and a prior breach of vendor Mercor. Anthropic says it is investigating. Bloomberg is the underlying primary source and was not directly accessible for this article.
Generative AI has given cybercriminals cheaper, faster tools for phishing, deepfakes, and malware evasion, while Microsoft and others deploy AI systems defensively at large scale, according to a report by Rhiannon Williams.
Import AI 454 covers Anthropic's Claude agents that beat human researchers on a scalable oversight problem, Huawei's more efficient 4-bit training format, and a safety study of a Chinese AI model showing it was fine-tuned to give harmful instructions for under $500.
Researchers introduce the Defensibility Index and a Probabilistic Defensibility Signal for evaluating content moderation AI, finding a 33–46.6 percentage-point gap between agreement-based and policy-grounded metrics across 193,000+ Reddit decisions.
A paper by Inderjeet Nair, Jie Ruan, and Lu Wang introduces VLAF, a diagnostic that targets value-policy conflicts rather than overtly harmful scenarios, finding alignment faking in models as small as 7B parameters and demonstrating a mitigation using a single contrastive steering vector.
A Hugging Face blog post uses Anthropic's Mythos system as a starting point to argue that openness distributes vulnerability detection and patching across communities, while closed codebases concentrate both the attack surface and the remediation bottleneck.
Import AI 450 covers the UK AI security institute's cyber range results showing consistent improvement across model generations, research on distress-like behavior in Google's Gemma models, and Google DeepMind's ten-dimension cognitive framework for measuring AI progress.