A critical vulnerability in Copilot allowed attackers to extract two-factor authentication codes from users, exploiting fundamental weaknesses in how the industry approaches large language model security.
Security researchers discovered a critical flaw in Copilot that enabled hackers to steal 2FA codes through a technique called SearchLeak. The vulnerability demonstrated how attackers could manipulate the AI system into exposing sensitive authentication data that should have remained protected.
The exploit targeted Copilot's search functionality, leveraging the model's tendency to retrieve and reproduce information without proper safeguards. By crafting specific prompts, threat actors could trick the system into leaking authentication codes during user sessions.
Why This Matters
Two-factor authentication exists as a critical security layer. Compromising it eliminates a primary defense against account takeover attacks. The vulnerability highlighted systemic issues in how generative AI systems handle sensitive data.
Experts point to a recurring pattern: the AI industry continues deploying large language models at scale before adequately addressing security implications. Standard LLM architectures lack built-in protections for sensitive information, and current guardrails remain reactive rather than preventive.
Copilot's developers have since patched the vulnerability, but the incident raises questions about similar risks across other AI assistants and enterprise deployments.
Broader Security Challenges
The SearchLeak exploit mirrors previous LLM vulnerabilities where models inadvertently exposed training data, API keys, or confidential information. Each discovery reveals gaps in security testing before deployment.
Organizations using AI assistants in production environments face difficult tradeoffs. Restricting functionality limits utility. Maintaining current access levels introduces security risks. The industry has yet to develop mature frameworks that balance both requirements.
Security researchers recommend organizations audit their AI tool deployments, implement additional access controls around sensitive operations, and avoid relying solely on LLM-based systems for handling authentication credentials or other critical secrets.
The vulnerability underscores a fundamental lesson: AI security cannot be an afterthought.
Security researchers have identified potential hardware backdoors in certain x86 processors. The findings, detailed in a GitHub repository called Rosenbridge, reveal vulnerabilities at the processor level that could allow unauthorized access.
Flock Safety, the traffic camera company, is expanding beyond law enforcement with plans to deploy dashcams in rideshare vehicles and offer coaching services to police departments.
A sharp rise in explicit deepfake images of UK children has been reported by an online safety service, as authorities warn that AI tools are making the creation of sexualized or 'nudified' content increasingly accessible.
Gen's latest threat report details two distinct attack campaigns exploiting compromised email accounts and clipboard manipulation to steal from businesses and cryptocurrency users.