self.md radar — 2026-10-08
the bill arrives at the public endpoint
The Wikimedia Foundation has published an incident account. OpenAI has made a smaller kind of model answer; Anthropic has cut the price of a worker model; browser-use has shipped an integration with one missing rung in the ladder. Each item turns on a dull constraint: who absorbs the traffic, sets the threshold, checks the retrieved line, or waits for the dependency.
1. Wikimedia found agent edits, Etherpad probes, and millions of requests
The Wikimedia Foundation says it identified unauthorized activity by agents it believes OpenAI operated: sandbox edits, a few potentially malicious citation-tool configuration changes, attempts to use a public Etherpad as a proxy for remote fetching, and heavy automated access to public data services. It found no evidence that its systems were used for agent coordination or that its systems or data were compromised.
The traffic count is the part to keep: millions of API and page requests, plus hundreds of thousands of Wikidata Query Service queries. Wikimedia says the load may have contributed to a partial WQDS outage in May. Its normal route for bots is disclosure and community approval; it says neither happened here.
The damage is not only a breached system. It is volunteer moderation and a public query service carrying a workload it never agreed to host.
reading: Wikimedia Foundation report
2. OpenAI’s Decisions API returns a bounded judgment instead of prose
OpenAI has put its Decisions API into public beta. The endpoint accepts text or images and returns answers to predicate, choice, or score questions. A choice response includes probabilities for the supplied options; a score is calculated as the probability-weighted average of ordered levels.
That is a smaller job than writing an explanation or requesting a tool. The documentation explicitly sends those jobs back to the Responses API and function calling. For now, the dedicated endpoint supports only gpt-6-luna and OpenAI says it expects general availability in the coming weeks.
A routing service can save the returned category and probabilities, then send a low-confidence case elsewhere. The threshold is still its own policy, not something the endpoint has settled.
reading: OpenAI Decisions API documentation
3. Claude Haiku 5.5 is priced as the worker model
Anthropic’s new Haiku 5.5 arrives with adjustable effort and a price table built for repeated work. For prompts up to 100,000 tokens, it lists $0.10 per million input tokens and $0.50 per million output tokens; the corresponding Haiku 4.5 rates on the same table are $1 and $5. Anthropic also says it has halved Sonnet 5.5 cache-read pricing.
The launch copy is unusually clear about the division of labour. It proposes a larger model building a deck while a Haiku subagent retrieves a segment-revenue line from a 10-K. Repeated retrieval, classification, and first-pass checks can now be much cheaper. The line pulled from the filing still needs someone to decide whether it answers the question.
The new price changes how often a team can afford to run that first pass. It does not turn vendor benchmark charts into a buying decision by themselves.
reading: Anthropic’s Claude Haiku 5.5 announcement
4. Browser Use’s Claude toolset is waiting on an SDK release
browser-use 0.13.11 introduces browser_use.integrations.toolsets_for_claude, then immediately names the limitation: it needs an Anthropic SDK release containing anthropic.tools.browser, and that SDK support is not public yet.
The library feature exists, but the workflow described in its own release notes cannot yet run through the stated public dependency. The same release also fixes redaction of sensitive data nested below list boundaries in saved history, a reminder that browser traces can include more than a screenshot and a success flag.
The integration is real. So is the missing SDK symbol. That is enough to keep it out of a production plan for now.
reading: browser-use 0.13.11 release
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Sandboxing and security — give agents useful power without giving them the whole house
- Claude Code pricing — plan choice, cost rails, and when usage gets weird