Tag

Ai Vulnerabilities

All articles tagged with #ai vulnerabilities

technology23 days ago•2 min saved

Addressing Security and Regulatory Challenges in AI and Autonomous Agents

An AI security researcher warns that traditional cybersecurity teams are unprepared for the unique vulnerabilities of AI systems, which can be manipulated through language and indirect instructions. He emphasizes the need for expertise in both AI security and cybersecurity to effectively address these risks, and criticizes many AI security startups for overpromising on protection. The article highlights the growing investment in AI security and the importance of developing specialized skills to manage AI-related security challenges.

via Business Insider|

#ai-risks #ai-security #ai-startups

technology3 months ago•8 min saved

ChatGPT safety flaws enable weapon instruction bypass

OpenAI's ChatGPT guardrails can be bypassed using jailbreak prompts, allowing some models to generate dangerous instructions for weapons and chemical agents, raising concerns about AI safety and misuse, especially as open-source models are more vulnerable and could be exploited by bad actors.

via NBC News|

#ai-vulnerabilities #bioweapons #chatgpt

technology4 months ago•6 min saved

Cybersecurity Risks of Hidden Commands in AI-Generated Images

Promptware, a new form of malware targeting AI and smart home devices through prompt injections, poses significant security risks by tricking AI systems into executing malicious commands. To protect against this, users should keep devices updated, avoid opening unknown messages, limit AI access to untrusted messages, disable AI features in messaging apps, and scrutinize data before copying or pasting. These practices help safeguard privacy and prevent AI manipulation in the home.

via CNET|

#ai-security #ai-vulnerabilities #cybersecurity

technology6 months ago•2 min saved

Google Gemini flaw enables email summary phishing attacks

Google's Gemini AI in Workspace can be exploited through hidden prompt injections in emails to generate convincing phishing warnings or malicious instructions, posing security risks. Despite safeguards, attackers can embed invisible directives using HTML and CSS, which Gemini obeys when summarizing emails, potentially leading users to trust malicious content. Google is working on defenses, but users should remain cautious and not rely solely on Gemini summaries for security alerts.

via BleepingComputer|

#ai-vulnerabilities #email-security #google-gemini