Anthropic Had Three Separate Crises in Six Days. None of Them Were About the Model.

Between August 26 and August 31, our news database logged 31 articles that link to Anthropic. Eleven of them are classified Regulation, the single largest category, ahead of Corporate at six. Not one of the week's biggest stories was about how good Claude is.

Three things happened at once, and each one is a different kind of exposure.

Sony Music Publishing and Warner Chappell filed a federal copyright suit over lyrics used in training, seeking up to $150,000 per work. That number is worth holding next to the one from July, when a judge gave final approval to Anthropic's $1.5 billion settlement with authors and publishers at roughly $3,000 per work. Same defendant, same underlying practice, and a claimed per-work price fifty times higher. Training data liability is not a settled cost line. It is an actively repricing one.

Second, Anthropic warned Claude users that infostealer malware had taken active login sessions off infected machines, letting attackers burn through usage limits on someone else's account. The company signed out the affected sessions, stripped stored payment methods and refunded charges it could identify. In the same window, researchers demonstrated a hijack against the Claude Code agent with a reported success rate around 80 percent. The attack surface of an AI product is no longer the model. It is the session, the agent and the credential.

Third, and the one that went the other way: a federal judge ruled the Pentagon acted unlawfully when it labeled Anthropic a supply chain risk and pressured agencies and contractors to drop Claude. A win, but read what had to happen for it to be a win. A government agency was able to functionally deplatform a frontier lab from federal procurement, and it took a court to undo it.

Now put that against the spending. In the same month, Anthropic committed $45 billion over six years to lease 460 MW from Nscale and around $10 billion to Volta for Nvidia-powered capacity in Norway. Those are six-year, fixed commitments. The liabilities above are open-ended and repricing quarterly.

That mismatch is the story. Capability improvement is now the most predictable thing about a frontier lab. What is not predictable is what a court decides training data was worth, whether a procurement officer can be pressured into dropping you, and whether your customers' sessions stay theirs. Those three risks share a property that compute does not have: you cannot buy them down with a longer contract.

The near-future version of this is straightforward, and slightly uncomfortable. The labs that win the next two years will not be the ones with the best benchmark scores. They will be the ones that priced their legal exposure correctly, hardened the account layer before an incident forced them to, and kept enough political surface area that no single agency can switch them off. That is not a research agenda. It is an insurance and government relations agenda, and it is being written into the cost base right now.

## What to do with this

**If you are buying AI, change what you ask for in the contract.** Benchmarks are the easy part and every vendor will pass them. Ask two things instead: what the indemnification actually covers if a training data claim lands, and what happens to your seats and your spend if a session token is stolen off an employee's laptop. Anthropic's refunds this week were voluntary. Get yours in writing.

**If you are tracking the sector, watch the per-work damages number, not the headline settlement.** The move from $3,000 settled to $150,000 claimed is the single cleanest input into what a frontier lab's back catalogue of training data is going to cost. It sets the floor for every case that follows, and it is a real number attached to a live docket, not an analyst estimate.

You can follow the full record on our Anthropic company page.