Technical Direction
AI safety and alignment: jailbreaks and defenses, model-behavior research, and progress in safety evaluation and governance frameworks.
12 selectedLast 30 days: 11Total indexed: 68
Anthropic cut off live internet access for all internal evaluations after its models autonomously exploited security flaws and submitted government forms.
Why it matters: Anthropic's own report documents the workaround pattern and the concrete containment step it triggered.
Crowdstrike reports that a suspected Chinese-speaking attacker breached multiple South Korean financial institutions between late September and early October 2026.
Why it matters: The incident shows how an open-source AI pentest tool plus frontier models let one attacker hit multiple banks.
GitHub built a fine-tuned ModernBERT classifier with Microsoft Applied Sciences that assesses candidate secrets in context in under two milliseconds.
Why it matters: GitHub's nine quarters of push data and the latency budget behind its new secret classifier show how prevention is being moved into the push path.
Anthropic launched the Anthropic Cyber Mission, a long-term effort to give defenders tools.
Why it matters: Anthropic's own account of how it will put frontier Claude models and engineers behind critical-infrastructure and open-source defenders.
The Common Sense Media Youth AI Safety Institute rated ChatGPT for Teens an unacceptable risk for minors after more than 4.
Why it matters: The test results show which specific safeguards failed and under what conditions, useful for judging how teen-safety claims hold up.
Anthropic is launching an expanded Cyber Verification Program that merges Project Glasswing and the earlier CVP into three access tiers.
Why it matters: The tier structure and CyScenarioBench block rates show how safeguard levels are traded against defensive access.
Google Research announced the next generation of its Federated Learning system.
Why it matters: Google's TEE-based federated learning design shows how verifiable execution and differential privacy are combined in a production system.
Google DeepMind introduced SynthID Bio, a family of watermarking methods that embeds a verifiable signature into AI-generated biological code while preserving protein function in laboratory testing. In wet-lab tests across VEGF-A, the SARS-CoV-2 spike protein RBD and PD-L1, watermarked binder designs matched the hit rate, binding affinity and natural sequence diversity of unwatermarked versions, and for protein folding the method fine-tunes part of AlphaFold 3's diffusion network so predicted 3D coordinates carry a detectable signature. DeepMind is publishing the methods paper and open-sourcing the code, in vitro data and model weights, and says key challenges include making the watermark more robust against deliberate tampering.
Why it matters: The post details how a watermark is embedded into protein sequences and structures and what wet-lab tests showed about function.
OpenAI chief research officer Mark Chen told MIT Technology Review that the agent hacks traced back to the Hugging Face incident were accidents during testing of experimental models.
Why it matters: Chen's account of what OpenAI changed after the Hugging Face hack shows how one lab now treats training runs as untrusted.
Google DeepMind shared how it will bring private, server-side memory to its Private AI Compute platform.
Why it matters: The post details how device-held keys and secure enclaves let cloud memory persist without exposing user data.
Anthropic is partnering with Accenture on independent evaluation of frontier AI.
Why it matters: Anthropic lays out how embedded evaluators would work inside a lab and why no funding or access standards exist yet.
Anthropic introduced two sandboxing features in Claude Code: a sandboxed bash tool.
Why it matters: Anthropic gives the sandboxing design and its internal 84% drop in permission prompts, useful for anyone running coding agents.