AI Engineering Weekly #9
This is the first model since GPT-2 that Anthropic has assessed as too dangerous for general release. The system card and red-team assessment are publ
Signals
This is the first model since GPT-2 that Anthropic has assessed as too dangerous for general release. The system card and red-team assessment are public
read both before forming opinions on what "step change" actually means here.
Web
Anthropic's Claude Mythos Preview
a model they're explicitly not releasing to the public — demonstrated autonomous zero-day discovery across major OS and browser targets, prompting Project Glasswing to gate access exclusively to vetted security researchers.
Web
Gemma 4 31B GGUF quant rankings by KL divergence are out
if you're running local inference, this is the fastest way to pick a quant without burning GPU hours on blind testing.
Web
GLM-5.1 targets long-horizon tasks
worth watching as a benchmark pressure point on frontier models for multi-step agentic workloads.
Simon Willison
Google open-sourced Scion, an experimental agent orchestration testbed
early-stage but signals Google treating multi-agent eval infrastructure as a shareable primitive rather than internal tooling.
Web
Japan relaxed privacy laws explicitly to become the easiest country to develop AI
regulatory arbitrage is now a stated national strategy, which affects where teams route data pipelines and fine-tuning workloads.
Web
MemPalace self-benchmarking exposed
a Reddit teardown shows the project claimed perfect scores on benchmarks it designed itself, a reminder that memory system evals are still almost entirely untrustworthy.
The Take
Mythos confirms that capability jumps in offensive security are now outpacing the defensive tooling and access-control frameworks teams have in place. Audit what your agentic pipelines can reach, because the threat model just moved.
Subscribe
Related Signals