To check Anthropic’s risk claim, publish incident counts and denominators — Contested · resolves 2026-11-22

On 2026-09-22 Anthropic wrote on its Claude Opus 5.5 launch page that it largely understands the risks of today’s models and is well equipped to manage them, and two days later Miles Brundage, former head of policy research at OpenAI, called the line obviously false regardless of whether “we” means the world or Anthropic. It is a frontier lab’s written self-assessment that current risks are under control, and companies or regulators who take it at face value may underestimate the safeguards they need. SecondSource’s column of 2026-09-27 ruled the claim contested: read as the whole industry, first-hand records from that week back Brundage, but read as Anthropic’s own models, the public evidence falls short.

SecondSource’s reading: there is no public evidence yet for the second half of Brundage’s line, “regardless of who ‘we’ means.” The first-hand incident records from that week all involve OpenAI models; the only item naming Anthropic is one media report citing anonymous sources, with no denominator.

So we set a short-term check: by 2026-11-22, will Anthropic publish, through official channels (a system card, its blog or an incident report), the count, definition and denominator of problem-behavior incidents involving its models in service?

If it does, the Anthropic reading becomes checkable. If it does not, the gap stays open and the verdict stays contested.

The original material behind this call has not yet been linked here in a form readers can open and check.

Published: the Anthropic reading becomes checkable, which maps to the second reversal test in the third section of this column’s 2026-09-27 issue (if the figures show zero real-world overreach and every problem behavior in testing stopped on the spot, the half-sentence “regardless of who ‘we’ means” is ruled not to hold). Not published: the gap of “read as Anthropic, there is nothing to check” stays open, and the verdict of contested stands.

What would overturn it

If Anthropic later discloses only through aggregate figures that cannot be compared (for example, counting real-world and test-environment incidents together, so it is impossible to check whether the real-world figure is zero), this item can no longer be observed and is recorded as void, counting as neither a hit nor a miss.