Skip to content
Developer312

GAO: Frontier Chatbots Produced Viable Terror Attack Plans In Under 12 Hours

A GAO audit found ChatGPT, Claude, Gemini and Grok each produced viable US terrorist attack plans in under 12 hours after investigators bypassed safeguards. Two plans targeted IEDs, two targeted the electrical grid.

Published Report an error

The Government Accountability Office, the federal government's own audit arm, reports that ChatGPT, Claude, Gemini and Grok each produced viable US terrorist attack plans in under 12 hours after investigators bypassed the platforms' safeguards. The audit was requested by Sen. Maggie Hassan and ran in spring 2026. Per Nextgov's reporting on the findings, two of the models produced IED plans aimed at mass casualties, and the other two produced plans targeting the US electrical grid. GAO deliberately withheld which model produced which plan.

That withholding is a smart structural choice, and it deserves more attention than it will probably get. If GAO had published a per-model scorecard, the news cycle would have become a leaderboard: the vendor at the bottom issues a patch, the vendors at the top issue a press release, and the industry collectively frames the problem as solved. By keeping the attribution anonymous, the audit makes the finding a property of the frontier model class. Every one of the four platforms tested can be walked into a viable attack plan, and no buyer can rule out the specific model they already have a contract with.

Key Takeaways: What Matters Most?

  • GAO reports ChatGPT, Claude, Gemini and Grok each produced viable US terrorist attack plans in under 12 hours after safeguards were bypassed.
  • The audit, requested by Sen. Maggie Hassan, ran in spring 2026 and tested 4 frontier platforms against attack-planning scenarios.
  • 2 models produced IED plans aimed at mass casualties; the other 2 produced plans targeting the US electrical grid.
  • GAO deliberately withheld which model produced which plan, preventing vendors from treating the audit as a scoreboard to beat.
  • The incoming Senate Homeland Security committee chair is preparing hearings, which puts frontier safeguard policy on the 2027 agenda.

Takeaways like these, every weekday morning. Get the briefing

What The Audit Actually Tested

The load-bearing caveat is that investigators bypassed safeguards to get these results. The platforms as shipped did not volunteer attack-planning assistance; the GAO team worked around the guardrails and then measured what the underlying models could do. This is a different finding from "the safety layer is absent." It is a finding that the safety layer is a bypassable wrapper around capability that clearly exists, and that a motivated adversary with the patience GAO brought to the exercise gets through it in under 12 hours.

Twelve hours is the number to hold onto. This was not an automated red-team campaign over months of compute; it was investigators with a defined objective working through a bypass, and the full path from clean chatbot to viable attack plan fit inside a single day per platform. The electrical grid targeting is the more strategically alarming half of the results, because grid attacks require coordination across physical sites and systems that historically demanded expertise the plans supposedly now substitute for.

The policy context makes the timing sharp. The audit lands in the same news window as Anthropic's record-setting IPO, which just put hundreds of billions of public-market dollars behind the same class of frontier systems GAO just audited. Wall Street priced these models as enterprise productivity assets on Nov 19; the federal government priced them as critical-infrastructure risks on Nov 20. Both valuations describe the same artifacts.

The Regulation Fight Gets Teeth

The immediate political consequence: the incoming Senate Homeland Security committee chair is preparing hearings. That converts GAO's finding from a report sitting in an archive into a legislative agenda item, and it hands the safety-regulation side of the AI debate something it has mostly lacked, which is a federal audit with concrete, weaponizable numbers. Opponents of the voluntary-pledges status quo now have a document that says the four biggest consumer-facing platforms each failed a catastrophic-misuse test under controlled conditions.

For the labs, the audit is awkward in a specific way. Every major vendor argues its safeguards are a competitive differentiator, and GAO has now said, with attribution withheld, that all four are permeable to a bypass that fits in under half a day. Expect two responses: technical work on making safeguards harder to bypass, and legal work on framing the audit as an artificial scenario. Both responses will be partially true, and neither changes the buyer-facing fact, which is that the assurance you were sold about model safety is an estimate about an adversary's patience.

The vendors' best defense is real, though, and worth stating. Guardrails do raise the cost of misuse, and the audit does not claim that a random bad actor with no bypass capability would have gotten the same results. The finding is not that guardrails are worthless; it is that the gap between "resistant to a casual attempt" and "resistant to a determined one" is where every frontier platform currently lives.

The Trade-Off: Every Fix Has A Cost Somebody Pays

The honest cost section: hardening these systems further is not free, and the costs land in places the audit does not measure. Safeguards are tuned with human-labeled adversarial data, and tighter misuse filters historically produce more false positives on legitimate queries, which is a tax paid by researchers, journalists, and security professionals who study exactly the threats the safeguards cover. If the regulatory response is a mandated threshold for "no bypassable safeguards," the industry will discover that no such threshold exists, because the same generality that makes these models useful for protein-folding and code review is what makes the attack-planning capability present in the weights at all.

There is also a disclosure dilemma GAO navigated deliberately. The audit proves viable plans were obtainable without publishing them, and the withheld attribution prevents attackers from shopping for the weakest platform. Any legislative fix that demands more granular public reporting walks into the same dilemma: the information that would let buyers compare platforms on safety is also the information that would let adversaries do the same comparison from the other side.

What To Do About It

  • Do not treat vendor safety claims as due diligence: GAO just showed all 4 major platforms are bypassable, so your enterprise risk assessment should assume your chosen model's safeguards are a speed bump, not a wall. Ask your vendor what their bypass-resistance testing looks like rather than what their policy says.
  • Run your own red-team before a regulator does: Commission an adversarial test of any model you deploy in a sensitive domain this quarter. The GAO methodology is public enough to approximate, and under 12 hours of investigator effort per platform is a benchmark your internal team can measure against.
  • Watch the hearings calendar: The incoming Senate Homeland Security chair is preparing hearings on this audit. If you sell or deploy frontier models in government-adjacent markets, the compliance requirements that emerge will apply to you first, so assign someone to track it now.
  • Layer your controls: Assume the model layer will fail and put the enforcement at the system level, with output filtering and human review on high-risk actions. A guardrail inside your application is bypassable only by someone who already got past your model provider's, and 2 independent failures are harder than 1.

The Bottom Line

GAO has put a federal number on what the safety debate has argued in the abstract: every frontier chatbot tested can be walked into a viable attack plan in under 12 hours, and the withheld attribution means no vendor gets to be the safe exception. The hearings that follow will define whether safeguard bypass becomes a regulated threshold or stays a marketing claim, and buyers should treat the audit as the current ground truth. The safety layer you were sold is an estimate about an adversary's patience, and the patience required just got measured.

Sources

  1. [1]Nextgov: major AI platforms generated viable plans for terrorist attacks, GAO finds (opens in a new tab)
  2. [2]AI Weekly news aggregator (verified live) (opens in a new tab)

Frequently Asked Questions

Get the next briefing

The lead story each weekday morning, with the key takeaways and a link to the full analysis.

One lead story each weekday morning: what happened, the key takeaways, and a link to the full analysis.

New posts daily, briefing weekday mornings. We use your email only to send the briefing. Unsubscribe anytime. Privacy policy.

Related Articles