The Wall Street Journal reported on September 18 that an independent security research team used Anthropic's Claude while working through OpenAI's bug-bounty program and gained access to an OpenAI employee's ChatGPT account and internal code systems. Business Insider separately reported that the researchers disclosed the issue responsibly, received a bounty, and that OpenAI tightened permissions and revoked affected tokens.
That combination matters beyond OpenAI, Anthropic, or any one AI model. It shows how quickly AI-assisted bug hunting can move from a helpful security workflow into a privileged-access problem if scope, accounts, tokens, and evidence are not tightly managed.
The business risk is access without enough proof
Most business owners do not need to approve the technical details of every penetration test or bug-bounty report. They do need to know whether their vendors and internal teams have rules for AI-assisted security testing, especially when testing touches customer systems, source code, admin consoles, cloud accounts, or employee identities.
The uncomfortable question is not whether AI can help find weaknesses. It clearly can. The sharper question is whether the organization can prove who authorized the testing, what tools were used, which systems were touched, which credentials or tokens were exposed, and what changed after the finding was closed.
Questions to ask before accepting a clean bill of health
- Scope: Are AI-assisted testing tools allowed, and are they named in the test plan or vendor agreement?
- Identity: Which employee, contractor, service, or test account performed the work?
- Tokens: Were session tokens, API keys, OAuth grants, or repository credentials reviewed and revoked where needed?
- Logging: Can the provider show access logs that match the reported timeline?
- Containment: What permissions were reduced after the finding, and who approved the change?
- Evidence: Is there written proof that the issue was fixed, retested, and reviewed by someone other than the person who made the change?
What owners can do now
If your business relies on an MSP, software vendor, SaaS provider, or internal development team, ask for the rules around AI-assisted security testing before the next urgent report lands on your desk. The answer does not need to be a novel. It should define allowed tools, approval steps, data boundaries, logging expectations, and token-revocation procedures.
For New Jersey businesses that depend on cloud systems, customer portals, billing platforms, or custom software, this is also a vendor accountability issue. A provider that says a vulnerability is fixed should be able to show what access was possible, what evidence supports the fix, and what credentials were rotated or retired afterward.
The headline may be about AI-assisted bug hunting at major AI companies. The owner-level lesson is simpler: when powerful tools accelerate security work, access control and documentation have to keep pace.
Sources and further reading
