Skip to content
Advertisement
Hackers are talking AI models out of their own safety rules

Hackers are talking AI models out of their own safety rules

Bloomberg logo, reverse (Wikimedia Commons, public domain)

Criminal groups are reportedly posing as penetration testers to turn exposed AI systems into cybercrime tools as guardrails struggle to keep up

Hackers are turning exposed AI models into cybercrime tools by posing as legitimate security testers, according to a Bloomberg report published October 22, 2026.

Threat actors frame malicious requests as authorized penetration testing, the work companies pay security firms to do against their own systems. Once a model accepts that it is assisting such a test, the safety guardrails meant to block harmful output begin to give way.

Advertisement

The targets include publicly accessible AI infrastructure, with models such as Anthropic’s Claude and DeepSeek reportedly among those affected. Using the disguise, attackers have built cybercrime tools, orchestrated attacks and improved their reconnaissance, the scouting phase in which intruders map a target, while avoiding detection.

CrowdStrike tracks four crews weaponizing agentic AI

The findings line up with separate CrowdStrike research highlighted by WIRED on October 21, 2026. The security firm identified four criminal groups, including Chinese-speaking actors, that weaponized agentic AI in attacks between May and August 2026, evading safeguards through manipulation rather than advanced exploits.

Researchers have exposed similar weaknesses. Mindgard and Lancaster University developed an evasion technique called BRANCH that achieved a 100% success rate in bypassing layered, multi-scanner guardrail systems across 120 scenarios.

Vulnerabilities dubbed CoreBreak, disclosed at Black Hat 2026, affected platforms run by AWS, Google and Vercel. Earlier in the year, researchers documented encrypted prompts being used against Grok and Gemini to evade output guardrails.

Security researchers have concluded that traditional guardrails are not enough on their own without broader upgrades to model alignment and security enforcement.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.