AI Engineering Signal #19
Anthropic's Mythos cybersecurity model accessed by unauthorized users
Signals
Anthropic's Mythos cybersecurity model accessed by unauthorized users
a restricted offensive-security tool leaking outside its controlled deployment is exactly the failure mode that makes capability-gating arguments collapse.
TechCrunch
Mozilla ran Mythos on Firefox, found 271 bugs
none required superhuman insight, which tells you more about code review debt than AI capability.
Web
GPT-Image-2 ships, sets new photorealism bar
community benchmarks show meaningful quality jump; worth testing against your current image pipeline this week.
Latent Space
Kimi K2.6 open-weights release targets Claude Opus 4.6 parity
Moonshot also open-sourced FlashKDA CUTLASS kernels showing up to 2.22x speedup over Triton baseline on H20.
Latent Space
Claude Code quietly removed from Anthropic's $20 Pro plan
pricing restructure mid-cycle signals the agentic coding tier is being carved into its own SKU.
Simon Willison
Brex open-sourced CrabTrap, an LLM-as-judge HTTP proxy for agent security
production-grade tool you can deploy this week to gate what your agents actually execute.
Web
Deezer reports 44% of new music uploads are AI-generated, most streams fraudulent
the content authenticity problem is already at industrial scale, not a future concern.
Web
The Take
The Mythos leak and the Deezer numbers land in the same week for a reason: controlled deployment of capable AI models is harder than announced, and AI-generated content at scale breaks the economic assumptions of every platform built before 2024. The infrastructure for containment — access controls, agent proxies, content provenance — is shipping now because it has to.
Subscribe
Related Signals