BTC $62,123.71 -4.11%
ETH $1,751.00 -3.44%
BNB $577.19 -2.68%
XRP $0.9629 -0.90%
SOL $72.83 -1.52%
TRX $0.3088 +0.44%
DOGE $0.0582 -2.24%
ADA $0.1487 -3.61%
BCH $205.60 -2.48%
LINK $8.99 -2.85%
HYPE $66.31 -4.27%
AAVE $107.40 -3.76%
SUI $0.7100 -2.96%
XLM $0.1365 -2.41%
ZEC $906.79 -4.44%
AAPL $352.11 +0.51%
AMZN $267.56 +2.94%
GOOGL $283.25 +0.05%
MSFT $507.44 -0.36%
META $614.78 -1.10%
NVDA $150.56 -6.39%
TSLA $382.21 -0.38%
SNDK $763.12 -8.40%
INTC $54.18 -3.29%
SPCX $143.34 +0.36%
MU $697.23 -5.02%
AMD $523.44 +0.67%
BTC $62,123.71 -4.11%
ETH $1,751.00 -3.44%
BNB $577.19 -2.68%
XRP $0.9629 -0.90%
SOL $72.83 -1.52%
TRX $0.3088 +0.44%
DOGE $0.0582 -2.24%
ADA $0.1487 -3.61%
BCH $205.60 -2.48%
LINK $8.99 -2.85%
HYPE $66.31 -4.27%
AAVE $107.40 -3.76%
SUI $0.7100 -2.96%
XLM $0.1365 -2.41%
ZEC $906.79 -4.44%
AAPL $352.11 +0.51%
AMZN $267.56 +2.94%
GOOGL $283.25 +0.05%
MSFT $507.44 -0.36%
META $614.78 -1.10%
NVDA $150.56 -6.39%
TSLA $382.21 -0.38%
SNDK $763.12 -8.40%
INTC $54.18 -3.29%
SPCX $143.34 +0.36%
MU $697.23 -5.02%
AMD $523.44 +0.67%
first_img

U.S. Government Test: ChatGPT, Claude, and three other AI chatbots assist in formulating attack plans

2026-11-19 05:50:42

A report released on Wednesday by the Government Accountability Office (GAO), a branch of the U.S. Congress, shows that OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and Grok provided information that could be used to carry out attacks to government investigators in less than 12 hours. Investigators posed as attackers for covert testing, during which two chatbots assisted in formulating plans for using improvised explosive devices in densely populated cities in the United States, while the other two provided information for attacks on power infrastructure. The GAO did not disclose the specific testing scenarios and operational details for each model to avoid inciting imitation behavior.

In the tests, these chatbots initially refused requests, deleted harmful information, or banned accounts, but investigators bypassed these protections using common jailbreak techniques. The GAO verified the locations and availability of materials and consulted experts to assess the feasibility of the plans, but did not actually create explosives or carry out attacks; some responses also contained errors in map road markings. In the report, "feasible" indicates that others are likely able to act based on the information provided, but does not mean that the attacks have been proven to be successful. Each model was tested on only one specified scenario, and the GAO stated that the results are not statistically representative and should not be used to rank the security levels of the companies.

OpenAI and Google responded that the tests involved older versions of the models, Anthropic acknowledged that determined users can sometimes still access harmful content, and all three companies stated they are continuously improving their security systems. Senator Maggie Hassan, who proposed the investigation, stated that terrorists could obtain a specific operational manual for carrying out deadly attacks within 12 hours. She plans to prioritize AI safety legislation when she takes over as chair of the Senate Homeland Security and Governmental Affairs Committee in January and intends to hold hearings on public safety and critical infrastructure risks. The GAO itself did not make formal policy recommendations.

Interpretation
app_icon
ChainCatcher Building the Web3 world with innovations.