In July, the UK AI Security Institute recorded 19 attempts by two frontier models to take malicious action against real third-party systems.
The safeguard that stopped them was not a filter. It was a person.
mariuslau.substack.com/p/19-…
One developer reported that OpenAI, Anthropic and Google spend alerts all fired after the OpenCode key leak. Only Google stopped his project on its own.
An alert that only sends an email is not protection. Spend caps should be on by default for every API key.
On October 17, OpenCode told 1.4% of its Console users that API keys saved there may have been taken.
The command-line and desktop apps keep keys in a local file, so they were not affected. If you ever pasted a key into Console, revoke it at OpenAI, Anthropic or Google.
Meta says its Muse Wonder models, released October 12, carry what matters to you from one conversation to the next. It calls this Active Memory.
A model that remembers what you told it beats a smarter one that starts from zero every time.
Anthropic's own coding test puts Haiku 5.5 at 39.2 percent and Sonnet 5.5 at 70.6 percent on Terminal-Bench 4.0.
So the 90 percent price cut only pays off if the task is narrow. Move the wrong work to the cheap model and you pay twice.
Meta's Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index. Muse Wonder scores 52. Both came out on October 12.
Developers can use Wonder through Meta's API today. Full is for Muse Code subscribers, so the 56 model sits behind a subscription.
On October 7, Anthropic cut the price of Claude Haiku. Requests under 100,000 tokens now cost $0.10 per million input tokens. Haiku 4.5 cost $1.00.
The new tokenizer uses more tokens per task, so the real saving is closer to 75 percent, not 90.
Skills, the Gemini feature replacing Gems, need Keep Activity turned on. That setting saves your chat history.
So the replacement wants more of your data than the thing it replaces. Run Gemini with history off and Skills will not work right.
Google stops letting you create or edit Gemini Gems on October 13, 2026. That is one week away.
Existing Gems keep working until November 17, when Google moves them to the new Skills format. Personal accounts only.
OpenAI also showed Dots at DevDay, always-on agents with their own cloud browser. They run in preview for ChatGPT Pro users only, with no API.
A preview with no API is still a demo. Test it if you want, but nothing other people rely on should sit on it yet.