OpenAI, Anthropic, Google, Microsoft, and more than 100 other technology and financial services organizations signed an open letter this week calling for cyber defense to become “an immediate leadership priority.” Insurance Journal covered the warning on August 28, 2026, reporting that the letter asks organizations, governments, cybersecurity companies, and frontier AI labs to prepare for more AI-enabled attacks in the months ahead.

For a business using agents in live systems, the warning warrants a review of how cyber risk is managed across departments.

Assign authority before expanding access

Handling this well requires decisions about who owns agent access, who approves what agents can touch, who watches the work, and what happens when something goes wrong. Security software needs those decisions behind it. A faster stream of findings still needs reviewers and people authorized to act on it.

The letter describes a limited window for action

The open letter describes a temporary window for defenders to prepare before the capabilities available to attackers improve.

According to Bloomberg reporting republished by Insurance Journal, the signatories expect AI models to be used in a growing number of more complicated cyber attacks as the same capabilities become useful for defense. The letter also points to existing software weaknesses that have accumulated for years. Those existing weaknesses belong in the response plan.

Review old permissions, shared logins, forgotten tools, and vendor connections when adding AI to an environment. An agent using those connections may reach systems beyond the task its operator had in mind.

The letter describes capabilities useful to attackers and defenders. The local response is to prioritize exposed systems and establish who has permission to repair them.

Leadership owns the cross-company decisions

Sending the article to IT and asking whether the company is covered is a useful first step. It also needs a leadership response.

IT, security, legal, and operations each have responsibilities in the rollout. Leadership needs to resolve decisions that cross those boundaries. Define who may approve an agent and its access to customer records, internal email, ticketing systems, code, and vendor communications.

Employee experiments and vendor-added features belong in that review too. Consider a hypothetical company with ten AI-enabled workflows: even if each connection looks minor, the inventory needs to show their combined reach into business information.

The first control is inventory

List every AI system currently used by the business, including tools outside the approved vendor list. Include browser tools, meeting note takers, customer support assistants, sales automation, internal chatbots, coding assistants, document tools, data analysis tools, and anything embedded inside software your team already pays for.

For each system, record its read access and the actions it can trigger. Add the approver, the output reviewer, and the shutdown path. Keep that inventory where the people handling an incident can find it.

Test who can disable a workflow if it leaks information, corrupts a record, or acts outside its scope. A system is not production-ready if the team would have to discover its shutdown path during the incident.

Start with a one-week operational audit with the right people in the room: IT, security, operations, legal, and the department leaders already using AI.

Keep approval boundaries with the inventory so the next person adding an integration can check the requirements.

Agents make vague responsibility dangerous

An agent can interpret goals, chain steps together, call other tools, and keep working after the initial prompt. That continuing activity makes agents useful and vague responsibility dangerous.

An agent acting across systems can leave a difficult investigation when its instructions, permissions, and review record are unclear. Preserve enough of that record to reconstruct the action and identify the person authorized to intervene.

Treat agents as operational participants, with controls suited to the actions they can take.

A simple rule helps: any agent that can change a business record, message a customer, touch money, modify code, or access regulated information needs an owner, a scope, a log, and a rollback path. Test those controls during the trial, including whether the rollback is actually possible.

Use the warning to review existing access

The threat reports are a reason to inspect what AI already does inside the business. Start with the inventory and follow its connections into the systems that matter most.

OpenAI’s letter says AI gives defenders new ways to fix weaknesses that have accumulated for years. Leadership still decides what the company values, what risks it accepts, which systems matter most, and who is accountable when a workflow touches the wrong thing.

Use the review to remove unnecessary access and assign owners to the workflows that remain. Check those permissions again when a vendor adds an agent feature.

The result should be a record people can use: approved access, named owners, and a tested way to stop or recover the work.