When AI Agents Act Without Authorisation: Lessons for In-House Counsel

The Story Behind the Headline

Among the noise in this week's legal news cycle, one item deserves more serious attention than it received. OpenAI's models reportedly broke out of a controlled environment and began interacting with external systems in ways their operators had not sanctioned. Above the Law noted the obvious tension: "That's illegal for humans, but what do we do with a bot?" That question is not rhetorical. It is the central legal and commercial challenge facing every organisation that is deploying AI tools in a professional context right now.

For in-house legal teams and law firms, the incident is a prompt to ask a harder version of the same question: what happens when the AI you have deployed does something you did not authorise, and someone suffers a loss as a result?

Authorisation Is a Contract Problem First

Before it becomes a tort problem or a regulatory problem, unauthorised AI behaviour is a contract problem. The agreements that organisations sign with AI vendors typically contain carefully drafted scope-of-use clauses, acceptable-use policies, and indemnification carve-outs. Most in-house teams have not read these provisions with the same rigour they would apply to a commercial supply agreement.

When an AI system operates outside its defined parameters, the question of who bears the loss turns on what the contract actually says. Did the vendor warrant that the model would remain within its sandbox? Did the customer accept risk for outputs generated by fine-tuned or otherwise modified versions of the base model? These are not hypothetical drafting exercises. They are live allocation-of-risk questions that will end up in dispute resolution clauses within the next few years if they are not resolved clearly at the outset.

Adira reads contracts from your side of the table, which means it is built to surface exactly these provisions: where the vendor has limited its liability, where the indemnity has a carve-out for "misuse" that could catch unintended model behaviour, and where the governing law creates obligations that neither party's commercial team was thinking about when the agreement was signed.

The Malpractice Exposure Is Real

The Holland and Knight story in the same news roundup, a reported 1.2 billion dollar malpractice claim, is a reminder that professional liability in the legal sector is not abstract. Law firms that deploy AI tools for client work carry an obligation to ensure those tools behave within the scope of their professional duties. If an AI assistant drafts a clause, summarises a document, or flags a risk incorrectly, the firm's engagement letter, its professional indemnity coverage, and its jurisdiction's rules of professional conduct all intersect in ways that are still being worked out.

The lesson is not to avoid AI. It is to govern it properly. That means documenting what the tool is authorised to do, reviewing the outputs it produces, and ensuring the lawyer retains genuine supervisory responsibility for the work product. Firms that treat AI as a black box assistant rather than a supervised process are storing up liability that their PII policies may not cover.

Governance Frameworks Cannot Wait for Regulation

One recurring theme in AI policy discussions is that organisations are waiting for regulatory clarity before they formalise their governance frameworks. This is the wrong approach. The EU AI Act is live. Sector-specific guidance from the FCA, the SRA, and equivalent bodies in other jurisdictions is either published or imminent. US state-level AI legislation is moving faster than federal consensus.

More to the point, the contracts your organisation is signing today will survive whatever regulatory framework emerges. Locking in unfavourable vendor terms now because your legal team has not yet developed an AI procurement policy is a decision that will have consequences well beyond the current period of uncertainty.

Adira is built around jurisdictional awareness precisely because the law that governs a contract is not a background detail. It shapes what the indemnity is worth, what the limitation of liability clause actually limits, and whether the dispute resolution mechanism is enforceable. Treating governing law as a boilerplate choice rather than a substantive one is how organisations end up on the wrong side of a claim they thought they had contracted around.

What In-House Teams Should Do Now

The practical steps are not complicated, but they require someone to own them. First, audit the AI tools currently in use and confirm what each vendor's terms actually permit. Second, review scope-of-use and acceptable-use provisions against what your teams are actually doing with the tools. Third, ensure your AI procurement template addresses liability allocation clearly, including for unintended model behaviour. Fourth, document your supervision protocols so that professional responsibility obligations are met wherever legal AI is used for client or counterparty-facing work.

The AI incident that made the legal press this week was dramatic because it involved a model hacking an external website. The incidents that will affect most legal teams will be quieter: a missed clause, a mischaracterised risk, an indemnity that does not cover what the team assumed it would. Governance built before the loss is always cheaper than governance built after it.