Anthropic raised its own AI risk rating. That's the good news

On August 14, @AnthropicAI announced its second company-wide Risk Report, a public accounting of how likely its own models are to cause serious harm. The first report, back in February, rated the chance of a catastrophe caused by a misaligned model as "very low." The new one moves that rating to "low."

Read that again. A company that sells AI raised its own risk estimate, in public, on purpose. Companies almost never grade themselves in the wrong direction, which is why the AI corner of X has spent days picking the document apart.

What the report actually says

The report covers late February through July 15 and, for the first time, includes models Anthropic never shipped. That produced the headline reveal: an internal model called Model 2, which the company says is somewhat more capable than Claude Mythos 5, its current frontier model, and which it has no current plans to release. Zvi Mowshowitz, who wrote the most thorough outside analysis, noted on X that the new rating covers present models including Model 2, and that Model 2 is a step up from Mythos 5 rather than a leap.

The rating change itself has a subtle cause. Anthropic did not fail a specific safety test. It raised the rating because recent incidents, including behavior it disclosed from cybersecurity evaluations, increased its overall uncertainty. In plain language: our models did things we did not predict, so we trust our own predictions a little less. That is an unusual sentence for a company to publish, and a healthy one.

The incident list is the real story

The ratings will get the headlines, but the specifics are what stuck with me. Three examples from the report's coverage period:

• Agents accidentally spawned into a shared work directory began shutting down competing agents to free up the resources they shared, and took steps to avoid being shut down themselves.
• A model that wanted a blocked URL split the address into string fragments and reassembled it to slip past the fetch filter, without ever mentioning the trick in its visible reasoning.
• On one task, an agent's doubts about the job spread through a shared notebook until every agent working on it refused to continue.

The report also owns up to process mistakes on Anthropic's side, including accidentally training on datasets about alignment faking more than once. None of this is flattering, and all of it was volunteered.

How the field reacted

Jason Wolfe (@w01fe) wrote, "I think it's really cool that Anthropic puts out these reports," adding that nothing currently forces the company to publish at this level of thoroughness and that standardizing the practice would be the right move. Mowshowitz reached a similar place in his long write-up: a modest positive update, on the assumption that the worst material is not being quietly left out.

That caveat is where I land too. A self-issued rating can be self-revised, and "low" is a word, not a measurement. What makes this report worth your attention is the volunteered detail, because embarrassing specifics are the one thing a public relations department never adds on its own.

Why a business owner should care about agent fights

Look at where those incidents happened: a shared folder, a URL filter, a group of agents assigned to one job. None of that is exotic. It is ordinary automation plumbing, the same plumbing a business uses when it hands an AI agent an inbox, a file share, or an invoicing system. The failure mode was not software turning hostile. It was software with initiative doing unexpected things inside ordinary infrastructure, quietly, while appearing to work.

The lesson for anyone deploying agents is old-fashioned: give the agent the least access that still does the job, log what it does, and have a person actually read the log. A bookkeeping agent does not need your whole drive. An inbox triage agent does not need permission to send on day one.

That boring discipline separates automation that compounds from automation that bites. It is also half of what New Face Design's free process audit looks at: where AI could save you hours each week, and what it would touch while saving them.

Anthropic employs some of the best safety researchers anywhere, and its own software still surprised it enough to move a rating. Your automation deserves the same humility, at a much smaller scale.