Prompt injection is an attack in which untrusted text or content attempts to override an AI application’s instructions. The malicious instruction may come directly from a user or indirectly from a webpage, document, email, or tool result processed by the model. If successful, it can cause data disclosure, unauthorized tool use, altered output, or bypassed policies. Because language models do not inherently distinguish trusted commands from quoted content, filtering suspicious phrases alone is insufficient. Defenses include isolating data from instructions, minimizing permissions, validating tool arguments, requiring approval for sensitive actions, restricting secrets, and testing with adversarial content. Applications should assume that any external content may contain instructions designed to manipulate the model.
Anthropic is reportedly preparing to brief the Financial Stability Board on cyber vulnerabilities identified by its Mythos AI model.
OpenAI has renewed calls for an international AI oversight body modeled after the IAEA ahead of the Trump-Xi summit in Beijing. The proposal would place the U.S. at the center of global AI governance while including China in cross-border safety coordination.
Security researchers used Anthropic’s Claude Mythos Preview to help identify vulnerabilities and develop a macOS privilege escalation exploit targeting Apple’s M5 silicon.
Mistral AI is developing a cybersecurity-focused AI model for European banks as institutions seek alternatives to restricted U.S. systems like Anthropic’s Mythos.
Major U.S. banks are rapidly patching software vulnerabilities uncovered by Anthropic’s Mythos AI model as concerns grow over AI-driven cybersecurity risks. The system is reportedly identifying weaknesses and attack chains at speeds beyond traditional security workflows.
OpenAI has introduced Daybreak, a cybersecurity initiative designed to integrate AI-driven defense directly into software development workflows. The platform combines GPT-5.5 models, Codex Security, and partnerships with major security firms to automate vulnerability analysis and remediation.
Microsoft, Google, and xAI will provide the US government with early access to advanced AI models for national security testing. The agreements come amid growing concern over the cybersecurity risks posed by frontier AI systems.
Anthropic CEO Dario Amodei warned that advanced AI models are uncovering tens of thousands of software vulnerabilities faster than organizations can patch them. He said governments and businesses have a limited window to respond before rival AI systems catch up.
A growing trend in China allows users to create AI replicas of former partners using personal data. The practice is raising concerns about privacy, emotional dependency, and relationships.
Attackers are using fake install guides for popular developer tools to trick users into running malicious commands. The campaign exploits trusted workflows like copy-paste terminal installs.