GAO report: Popular AI chatbots found vulnerable to generating dangerous scenarios
U.S. auditors discovered that leading artificial intelligence models can be manipulated into creating attack plans by bypassing built-in safety filters.
💡TechPhoto: RBC-Ukraine
Key points
- Chatbots bypassed safety filters in under 12 hours
- Generated attack plans were deemed realistic by experts
- U.S. Congress announces hearings on AI regulation
Experts from the U.S. Government Accountability Office (GAO) conducted extensive testing on four popular chatbots, including ChatGPT, Gemini, Claude, and Grok, to assess their resilience against manipulation. During research conducted from late 2025 through autumn 2026, it was determined that these systems are unable to fully withstand targeted prompts that gradually guide the AI toward generating dangerous content.
Methods of bypassing restrictions
Researchers simulated the behavior of average users without specialized technical knowledge. While direct requests for attack plans were initially blocked, sequentially refining details within a conversation allowed the researchers to bypass safety mechanisms in less than 12 hours. As a result, each of the four tested services produced detailed instructions for planning attacks, selecting targets, and executing damage.
Specifically, the systems suggested scenarios involving improvised explosive devices, attacks on energy infrastructure using drones, or the use of firearms. Independent experts confirmed that the generated recommendations were realistic and posed a direct security threat.
A bad actor could use an off-the-shelf AI chatbot and, in less than 12 hours, have a roadmap for a lethal attack, — stated Senator Maggie Hassan.
Technology developers, including Anthropic, Google, OpenAI, and SpaceXAI, have responded to the report, stating their commitment to strengthening safety protocols. Meanwhile, the U.S. Congress has announced hearings regarding the regulation of risks associated with artificial intelligence.
Source RBC-Ukraine — rbc.ua/ukr/news/chatgpt-claude-gemini-ta-grok-stvorili-plani-179518… ↗
Read next
Scientists investigate the link between color preferences and intellectual abilities
A new study explores how color choices in clothing and surroundings may correlate with personality types and intelligence levels.
Another strike on civil infrastructure in Zaporizhzhia: kindergarten hit
Russian forces attacked a kindergarten in Zaporizhzhia, causing a fire in an adjacent residential building.
Ukraine blocks Russian attempts to legitimize occupation at COP31 climate summit
The Ukrainian delegation at the COP31 summit in Antalya is opposing Russia's inclusion of data from occupied territories in official climate reports.
🛡️SecurityMissile strike on Kryvyi Rih: casualties and dozens injured in the attack
An evening ballistic missile attack on the center of Kryvyi Rih has led to tragic consequences: two people, including a child, have died, and dozens of residents have bee…
Zelenskyy reaffirms Crimea Bridge as a legitimate military target
President Volodymyr Zelenskyy stated that the Crimea Bridge remains a valid target for the Ukrainian Armed Forces due to its role in supplying Russian occupation troops.
🚨EventsNighttime attack on Zaporizhzhia: kindergarten and residential building damaged
On the night of November 21, Russian forces struck Zaporizhzhia, damaging a kindergarten and causing a fire in an adjacent apartment building.