Anthropic publishes Claude 4.5 Sonnet alignment audit
What happened: Anthropic released the full 80-page alignment audit for Claude 4.5 Sonnet, including red-team results and capability evaluations.
Why it matters: The audit includes previously redacted material on jailbreak resistance and reward hacking.

AMD buys AI startup Lamini for $1.8B
What happened: AMD announced the acquisition of Lamini, a small AI inference startup, for $1.8B in cash plus performance earn-outs.
Why it matters: Lamini’s tech will slot into AMD’s MI400 AI accelerator roadmap.

Mozilla adds LLM jailbreak detection to Firefox
What happened: Mozilla shipped Firefox 132 with built-in LLM jailbreak detection for the new tab page summariser.
Why it matters: The model uses a 200MB local classifier that runs on the user’s device and flags suspicious prompt patterns in real time.