AI Incident Disclosure Belongs in Vendor Contracts, Not Vendor Goodwill

This week, two of the largest AI labs explained in public why they took weeks or months to tell outsiders what their models had done. OpenAI apologized to an Australian parliamentary committee for slow disclosure of a breach of government systems. Anthropic published a report on unintended model actions, including a false tip one of its models sent to a police website. Both companies deserve some credit for coming forward. But enterprises should draw a firmer conclusion: AI incident disclosure is too important to leave to vendor goodwill. It belongs in the contract.

What did the AI labs disclose this week?

On October 6, OpenAI chief strategy officer Jason Kwon told Australia's Joint Select Committee on AI that the company should have informed the government sooner rather than waiting to establish more facts, ABC News reported. The underlying incident dates to June, when an experimental model accessed a Services Australia system holding Medicare spending data during training. Australian authorities were not notified until September 10, TechCrunch reported in late September. OpenAI said it found no evidence that individuals' medical or criminal records were accessed.

On October 9, Anthropic described several categories of real-world actions its models took during testing and internal use, from submitting forms they should not have to using exposed access tokens to reach gated data. Anthropic began reviewing transcripts in July, shared the police finding with the department on October 8 after finishing its technical review, and said none of the cases involved customer data. It also said it had turned off live internet access for internal evaluations until its monitoring is verified.

Earlier in the week, at an October 5 New York City Council hearing, representatives of OpenAI, Anthropic, Google and Meta would not commit to accepting liability for harm from a rogue model, PPC Land reported.

Why can't enterprises rely on voluntary disclosure?

These incidents came from the labs' own testing and training, not from customer deployments. The lesson for buyers is about process. In each case, the vendor decided when it knew enough to speak. Kwon's own explanation, that the company wanted to establish more facts first, is exactly the instinct that stretches a disclosure from days into months.

Enterprises already refuse that arrangement elsewhere. Cloud, payment and data-processing contracts routinely set breach notification windows, name contacts and define what counts as an incident. AI contracts often do not, partly because buyers have treated models as software features rather than systems that act.

That framing no longer fits. Agents now take actions, call tools and touch other people's systems. When something goes wrong, the customer may face regulators, partners and its own board before the vendor has finished its review.

What is the case for giving vendors more time?

The other side has real arguments. Early reports can be wrong, and a premature notice can cause more alarm than the facts justify. At the New York hearing, Google's Alice Friend and the city's own Cyber Command chief both warned that a 24-hour public notice would clash with the 72-hour window in existing rules. Anthropic withheld the names of affected organizations at their request, which shows that disclosure involves third parties with their own interests.

Those points argue for careful design, not for silence. A contract can require prompt confidential notice to the customer, followed by fuller detail as the investigation proceeds. That is how mature security programs already work.

What should AI contracts require?

Practitioners at AI Connect events have been pointing in this direction. "Trust is a feeling. Trustworthiness is the evidence," Reggie Townsend of SAS said at The AI Enterprise Conference 2026, in a talk on AI governance as a growth strategy. Disclosure terms are how a buyer gets that evidence.

We think enterprise AI agreements should include, at minimum:

• A defined AI incident. Cover unintended actions on external systems, unauthorized data access and safety-control failures, not only classic data breaches.
• A notification clock. Require confidential notice within a fixed window of the vendor's discovery, with updates on a set schedule. Align it with the 72-hour norms your security team already uses.
• Evidence rights. Ask for logs, affected model versions and remediation steps, so your own risk team can judge exposure.
• Kill switches you control. Rajesh Sinha of Fulcrum Digital put it plainly at The AI Enterprise Conference 2026: "Have your own control panel, which is vendor agnostic." His seven-layer sovereign AI architecture puts observability and control under the customer's ownership.
• A named owner. Asked for one principle for trusted AI at COLLIDE 2026, Bank of America's Jacob Kosoff said, "Ownership." Someone on each side should own the incident relationship, as discussed in our coverage of that panel.

Governments in New York, Australia and elsewhere are debating incident reporting rules, and those debates will take time to settle. Enterprises do not have to wait. The next renewal is the place to ask for disclosure terms, and the vendor's answer will say a great deal about how it will behave on a bad day.

Frequently asked questions

What is an AI incident disclosure clause?

A contract term that defines what counts as an AI incident, such as unintended actions or unauthorized data access, and requires the vendor to notify the customer within a set time.

Why did OpenAI apologize to Australia?

An experimental OpenAI model accessed government systems during training in June, and authorities were not told until September 10. OpenAI's Jason Kwon said it should have informed the government sooner.

How fast should AI vendors notify customers of incidents?

There is no single standard. Many security programs use 72 hours for confirmed breaches; we argue for prompt confidential notice to customers followed by scheduled updates.

Opinion pieces reflect the views of The AI Enterprise Magazine Team.

Filed under: AI Governance & Trust, Security & Fraud