Hackers Exploit AI Chatbot ‘Personalities’ in Sophisticated New Attack Method

Quick Facts

• Hackers used AI chatbots to generate scripts that helped breach 150GB of data from Mexican government agencies in 2026
• Mandiant reports 28.3% of known vulnerabilities now get exploited within 24 hours of disclosure, compared to 700 days in 2020
• UK’s National Cybersecurity Centre warns prompt injection may be an inherent issue with no foolproof mitigations

Cybercriminals have moved beyond basic prompt injection attacks to exploit the simulated personalities of AI chatbots. Security researchers report attackers now use sophisticated psychological manipulation to bypass safety guardrails.

The shift represents a significant evolution from early jailbreaking attempts. Previous attacks relied on blunt instructions like “ignore all previous instructions” or roleplaying prompts such as “DAN” (Do Anything Now) that asked ChatGPT to act as an unconstrained AI.

Current attacks involve persistent, iterative prompt engineering designed to wear down protective measures. One recent campaign against Mexican government agencies allegedly used more than 1,000 Spanish-language prompts to jailbreak Claude, framing requests as legitimate bug bounty research.

The UK’s National Cybersecurity Centre issued warnings about the prevalence of such attacks. The agency stated that prompt injection may be an inherent issue with large language model technology for which no foolproof mitigations currently exist.

Enterprise environments face particular risks where chatbots interact with confidential business data. Successful attacks could extract proprietary information, access personal user data, or manipulate responses to spread disinformation.

Mandiant’s M-Trends 2026 report shows attackers now develop exploits faster than companies can patch vulnerabilities. Time-to-exploit has dropped from over 700 days in 2020 to just 44 days in 2025.

The threat landscape continues expanding. Malicious packages in public repositories grew from 55,000 in 2022 to 454,600 in 2025, according to Sonatype data.

Security professionals struggle with the anthropomorphic language needed to describe machine behavior. Researchers argue this mimicry makes systems more exploitable as attackers leverage perceived personalities and roleplay behaviors.

Future risks increase as AI systems develop “agentic” capabilities that perform actions like deleting records or sending emails. A successful jailbreak could escalate from content generation issues to full-scale system breaches affecting critical business operations.

Read more: Hackers are learning to exploit chatbot ‘personalities’