AI Engineering Signal #45
Anthropic Mythos-class models confirmed for public release
Signals
Anthropic Mythos-class models confirmed for public release
security tooling teams should verify whether current API access routes to these models or needs separate provisioning.
Web
Memory now nearly two-thirds of AI chip component costs
inference cluster TCO models need immediate revision before next procurement cycle.
Web
LLM agents drift from constraints mid-task in backend code generation
add checkpoint validation gates before agents touch production code paths.
ArXiv
Inaudible audio hijacks AI voice assistants via hidden prompts
voice-interface deployments accepting ambient audio need adversarial input filtering before production.
Web
DeepMind agent solves 9 open Erdős problems autonomously
math-reasoning agents now warrant evaluation for formal verification and proof-assistant tooling.
Web
AMD RDNA3 gets fast native Qwen 3.6 inference via hipEngine
non-NVIDIA local inference has a viable routing option worth benchmarking this week.
The Take
The cost floor for inference is rising at the hardware layer while the attack surface is expanding at the input layer — teams that haven't audited chip procurement assumptions and agent input validation are carrying compounding risk on both fronts.
Subscribe
Related Signals