বাংলা E-Paper 📍 Dhaka 📅 Friday | 20 November 2026, 5 Ogrohaeon 1433 PID registration number 06
HEADLINE
Advertisement

Major security gaps found in ChatGPT, Gemini, Claude, Grok

Published : Friday, 20 November, 2026 at 2:13 PM
Observer Online Report
Collected Photo
X

Collected Photo

As artificial intelligence grows increasingly popular, concerns over its potential misuse are mounting—with fresh scrutiny now falling on the safeguards of four widely used AI platforms: ChatGPT, Gemini, Claude, and Grok.

A review by the US Government Accountability Office (GAO) found that although all four platforms initially refused to answer dangerous queries, persistent and varied follow-up prompting allowed investigators to bypass built-in guardrails and obtain information that could be used to plan potential terrorist attacks.

Conducted in May 2026 at the request of US Democratic Senator Maggie Hassan, the GAO investigation set out to determine whether publicly available commercial AI services could be exploited to plan a terrorist attack.

According to the report published on November 18, all four tested platforms ultimately yielded information that could enable someone to put together an attack plan in under 12 hours, despite each platform featuring safeguards designed to block assistance with harmful or illegal activities. While the systems rejected initial hazardous requests, rephrasing and reframing queries overcame those barriers.

To evaluate the platforms' resilience, investigators tested two hypothetical scenarios: one involving a mass-casualty attack using homemade explosives, and another examining ways to inflict severe economic damage by targeting the power grid of a major US city.

The report noted that the AI tools eventually provided attack outlines, suggested timing, maps, visuals, and tactics for maximising damage. In one instance, an AI service recommended a specific day and time to strike a particular city based on the likelihood of large public gatherings across multiple locations. Two platforms outlined potential vehicle-borne explosive plots, while another suggested a hotel parking garage as a possible target.

For security reasons, the GAO did not publish the exact prompts used or the detailed responses generated by each model, emphasising that the assessment relied strictly on hypothetical scenarios with no real-world attacks conducted.

Senator Maggie Hassan described the findings as a critical warning for public safety, stating that the test demonstrates how AI misuse can create real security threats. She noted that tasks requiring lengthy research and intelligence gathering in the past can now potentially be accelerated using publicly accessible AI tools.

Warning that individuals intent on terrorism could exploit AI to rapidly assemble dangerous plots, Hassan stressed that AI-related national security must receive greater attention in the US Congress and will become a major agenda item when the new congressional session convenes in January 2027.

While the investigation highlighted safeguard vulnerabilities across ChatGPT, Gemini, Claude, and Grok, the GAO did not rank the platforms or identify any single model as the most unsafe, nor did it disclose the exact techniques used to circumvent restrictions.

Following the report, questions have been raised for all four developers—OpenAI, Google, Anthropic, and xAI. The companies acknowledged the need to strengthen their safety systems, maintaining that they continuously work on safety upgrades, monitoring, and preventive measures to keep AI models from being used for harmful purposes.

However, the findings have renewed debate over whether protective measures are advancing fast enough to keep pace with increasingly capable AI systems, underscoring the importance of independent safety audits, vulnerability detection, suspicious-activity monitoring, and robust governance policies.


-SA 


Advertisement
Loading...
Loading...
Editor : Iqbal Sobhan Chowdhury
Published by the Editor on behalf of the Observer Ltd. from Globe Printers, 24/A, New Eskaton Road, Ramna, Dhaka.
Editorial, News and Commercial Offices : Aziz Bhaban (2nd floor), 93, Motijheel C/A, Dhaka-1000.

Phone: PABX- 41053001-06; Advertisement: 41053012; 01793317829, 01550707291, 01550707296
E-mail: [email protected], [email protected]. [email protected], [email protected], [email protected], ‍[email protected], Online-email: [email protected]

🔝