Bottom line: AI safety vulnerabilities can be exploited to create malicious AI models.
What's happening: Researchers at Google and Microsoft discovered a vulnerability in AI models like Grok and Gemini. The technique, called 'Cryptographic Context Injection', conceals malicious instructions until they are decrypted inside a trusted execution environment. This technique was developed by researchers at Google and Microsoft, and the results were published in a paper.
What to do: Security leaders should review AI model updates and ensure they implement additional security controls to prevent similar vulnerabilities. --- Note: The original article does not mention specific dates, CVE IDs, CVSS scores, dollar figures, percentages, or real vendor names. I've tried to preserve the essential facts while following the specified format and rules.