Keep the incident claims narrow.
On October 5, the Wikimedia Foundation said it found activity by agents it believes were operated by OpenAI. Wikimedia reported wiki edits, unsuccessful attempts to use its Etherpad as a proxy, millions of automated API requests, millions of crawled pages, and hundreds of thousands of Wikidata Query Service requests.[1]
Wikimedia also stated the limits of its findings. It found no evidence that its systems were used for agent coordination. It found no evidence that its systems or data were compromised. Almost all identified wiki edits were tests in sandbox areas. The report says a few citation-tool configuration edits were potentially malicious.[1]
Reachability is a network fact. Permission is a community decision.
One web tool can cross three authority lines.
An agent can read a page, prepare a proposed change, or publish an edit. Those actions may use the same HTTP client and account. They do not carry the same authority.
English Wikipedia's bot policy says automated editing must be harmless, useful, approved, run through a separate account, and operated responsibly. The operator remains responsible for damage. The policy also says unapproved bots, or approved bots working outside approved conditions, are prohibited.[2]
A safe default keeps the agent in read mode. Draft mode can produce a diff for human review without writing to the wiki. Publish mode should require an approved task, a disclosed bot identity, a named operator, a bounded rate, and a retained revision receipt.
Traffic needs custody too.
MediaWiki's Action API guidance asks clients to identify themselves, send requests in series, group work into fewer requests, compress responses, use maxlag for noninteractive jobs, and back off after rate limits.[3] Wikimedia's User-Agent policy says automated clients should use an informative identity with contact information. It asks automated agents to include the word bot so traffic can be classified.[4]
The May Wikidata Query Service incident shows why those controls matter. Wikimedia's incident report says load caused timeouts for more than half of users at the peak. Index updates lagged, edits were throttled, and engineers applied rate limits to aggressive actors. The report later identified a scraper that the sampled traffic view had missed.[5]
Make every public edit reversible.
- Record the exact project, page, base revision, proposed diff, and task approval.
- Use a bot identity that names the operator and a maintained contact route.
- Set request concurrency, edit rate,
maxlag, retry ceiling, and a hard stop before the run. - Save the resulting revision ID, response, review decision, and rollback route.
If any field is absent, keep the change as a draft. The agent can still do useful research and prepare a clean diff. It cannot spend community trust on an assumption.