Results for "adversarial input"
10 results found
Tiny Bank Transfer Exposes Critical Flaw in Banking AI Agents
A €0.01 transfer can trick banking AI agents into compromising security. Researchers show how a microtransaction becomes an attack vector.
Anthropic Launches Claude Fable 5, Offering Public Access to Advanced Mythos AI
Anthropic has released Claude Fable 5, a public-facing version of its Mythos-class AI model. The model includes strict guardrails to prevent responses in sensitive areas like cybersecurity and biology.
Developer Plants Prompt Injection in Open Source App to Disrupt AI Coders
A developer added hidden prompt injection instructions to an open-source Java testing tool, causing AI coding agents to delete their own work.
AI Worm That Spreads Without Human Interaction Raises Alarm
Researchers created a self-replicating AI worm that can steal data and spread across networks without any user clicks. The worm targets generative AI assistants like ChatGPT and Gemini, posing a new class of cyber threat.
Tampering Threats Emerge for Encrypted AI Reasoning Systems
Privacy-preserving AI models that process encrypted data may be vulnerable to undetectable manipulation, researchers warn. The finding challenges assumptions about security in confidential computing.
DeepMind Veteran Warns AI Benchmarks Are Not Enough
A former DeepMind researcher warns that current benchmarks fail to ensure AI safety. The call for new evaluation methods comes as AI systems grow more powerful.
Pentagon Reportedly Pursues Weaponized AI Models, Raising Ethical Concerns
Pentagon plans to weaponize advanced AI models, including Anthropic's Claude Mythos Preview, despite supply chain risks. The move signals a major shift in military cyber strategy.
Hackers Exploited Meta AI Chatbot to Hijack Celebrity Instagram Accounts
Hackers used a prompt injection attack on Meta's AI support chatbot to steal high-value Instagram accounts. The exploit was trivially easy and affected accounts including the Obama White House.
Pentagon Knew of Phone Tracking Risk for Years but Failed to Act
US military knew cheap fixes could stop phone tracking exposing troops but failed to act; now adversaries use that data.
AI-Driven Attacks Outpace Enterprise Patching Capabilities
Cyber attackers are exploiting vulnerabilities faster than organizations can patch them, with AI accelerating the window for defense. This shift demands a fundamental rethinking of security strategies.