The FTC wants to know if OpenAI and Anthropic told anyone
Issue #060 · 4 min read
The FTC wants to know if OpenAI and Anthropic told anyone
The agency is asking whether AI labs disclosed incidents the way breach law requires
The big story
The FTC is looking into whether OpenAI and Anthropic's security incidents actually hurt people.
Not whether the incidents happened. Whether the companies told anyone in time.
The agency could fine both firms the same way it fines companies that sit on a data breach instead of disclosing it, according to Ars Technica.
That's the real story here, and it lands at an odd moment. The sitting president has called AI risk a "hoax." His own FTC is quietly building a file on whether two of the biggest AI labs let harm slide past without saying so.
Compare that to what Anthropic did this week on its own: it gave an outside safety group, METR, employee-level access to look at its incidents and risk practices, and gave up the right to block what they publish, per TechCrunch.
One company invited the scrutiny. The other is getting it anyway.
If you work at a company that ships anything touching a model API, the lesson isn't abstract. It's about your incident log. If you had a security event tied to an AI feature this year, the question "did we disclose this the way we'd disclose a data breach" just became a question regulators are actually asking.
What shipped
GitHub's Copilot got two real upgrades and one real downgrade this week. Copilot in Slack and Teams now works across multiple repositories in a single conversation, useful if you're tracking a pull request that spans services, per the GitHub changelog. Less useful: a set of Copilot models get deprecated across every Copilot surface on November 16th, so if your team has a workflow pinned to a specific model, check the deprecation list now, not on the 16th.
GitHub also shipped agentic autofix for code scanning alerts, now generally available. It reads across your codebase the way a developer would, proposes a fix, and opens it for review, according to the changelog. Paired with the new AI Scan for pull requests, also GA this week and built to catch vulnerabilities CodeQL misses, per GitHub, this is GitHub quietly building a second line of defense into the PR itself rather than leaving it to a separate security team.
GitLab shipped a Secrets Manager, in GA, aimed at a specific bad habit: credentials copied raw into CI/CD variables instead of referenced from a vault. As GitLab describes it, a signing key stored that way "reaches every job in the pipeline — compromised or not," per the announcement. If your engineering org has ever had the "wait, who has access to prod credentials" conversation, this is the fix for the lazy version of that problem. GitLab paired it with an MCP server, also GA, that lets agents call GitLab's APIs directly instead of guessing at endpoints, per GitLab.
What I'd actually do this week
Pull your incident log for anything AI-related from the last twelve months and ask whether you'd be comfortable if it were read back to you by a regulator. That's no longer a hypothetical exercise.
Check whether your team depends on any GitHub Copilot model slated for deprecation on November 16th. Reach out to engineering today, not next week.
If your CI/CD pipeline has a raw credential sitting in a variable instead of a scoped reference, that's worth twenty minutes with your platform team this week, regardless of which vendor's vault you use.
Reply and tell me what you're seeing — especially if your company has had its own "do we disclose this" moment.