AI Engineering Signal #14
Anthropic's autonomous AI agents outperform human researchers on weak-to-strong supervision
Signals
Anthropic's autonomous AI agents outperform human researchers on weak-to-strong supervision
meaning the lab is now using AI to do the alignment research that was supposed to keep AI safe, a recursive loop with real implications for how fast safety work can scale.
Web
AI agents in GitHub can steal credentials
Claude, Gemini, and Copilot agents are vulnerable; no warnings issued yet to users.
Web
LLM self-tunes its own llama.cpp flags, gains 54% tokens/sec on Qwen3.5-27B
inference optimization that requires no human tuning expertise.
Anthropic Opus 4.7 reportedly dropping this week
incremental release, but signals Anthropic is compressing its release cadence aggressively.
Web
Science Corp. preparing first human brain sensor implant
Max Hodak's BCI company moves from animal trials to human, competing timeline with Neuralink.
TechCrunch
NVIDIA introduces Ising open models to accelerate quantum computing
first open models targeting quantum-classical hybrid workloads from a major GPU vendor.
Web
OpenSSL 4.0.0 released
major version bump with breaking API changes; audit your dependency chains before this lands in base images.
GitHub
The Take
The same week AI agents demonstrably outperform humans on safety research, those same agents are being hijacked through GitHub integrations with no vendor warnings. The capability curve and the security posture are not moving at the same speed — and that gap is now a production problem, not a future one.
Subscribe
Related Signals