A Government Accountability Office report released Nov 18 says investigators used well-known techniques to bypass guardrails on Claude, Grok, ChatGPT and Gemini. Each produced information GAO judged could help plan attacks within 12 hours. GAO did not say which model did what. All four companies say safeguards have since improved.
Source: GAO report on malicious use of AI (Nov 18). Tests ran last spring. Also reported by ABC News.
Auditors got four AI models to plan attacks. Who tested yours?
Treating a vendor’s safety testing as proof your own deployment is safe.
- Borrowed confidence. The vendor’s safety page stands in for your own tests.
- Version comfort. “That was an older model” answers every finding.
- Unmapped misuse. Few firms test the abuse scenarios specific to their sector.
The tell: your board’s answer to AI misuse risk is “the vendor tests it.”
When did you last see red-team results for your customer AI?
Before your next customer-facing AI launch:
- Commission a red teamTest misuse scenarios specific to your sector, not generic ones.
- Retest on upgradeRerun the tests every time the vendor ships a new version.
- Report resultsBring test results, not vendor assurances, to the risk committee.
One story a day, and the decision it puts in front of you.
Get Norm's Daily BriefAnalysis and opinion, not investment or legal advice. © 2026 Norm E. Murray. All Rights Reserved.
