Advertisement
Hiding Malicious Commands from AI via Font-Rendering Manipulation
Learn how attackers use font-rendering tricks to bypass AI safety filters and execute prompt injection attacks against LLM-powered assistants.
OpenClaw AI Agent Flaws: Prompt Injection and Data Exfiltration Risk
CNCERT warns of critical security flaws in OpenClaw AI agents, enabling prompt injection and data exfiltration due to weak default configurations.
Manipulating Perplexity Comet AI via Reasoning-Based Phishing
Researchers from Guardio demonstrate a rapid attack vector against Perplexity’s Comet AI browser, tricking it into executing malicious phishing tasks.
AI Agent Security Risks: Defending Against Autonomous Tool Misuse
Analysis of the security implications of autonomous AI agents, focusing on prompt injection, privilege escalation, and the erosion of trust boundaries.
AI Recommendation Poisoning: Manipulating Summarization Features
Discovery of hidden prompt injection in AI summarization buttons allows companies to bias AI memory and manipulate future user recommendations.
Anthropic Patches Claude Code Vulnerabilities Enabling Silent Hacking
Anthropic addressed flaws in Claude Code that allowed attackers to execute arbitrary commands on developer devices via malicious repository configurations.
Advertisement
RoguePilot Vulnerability: GitHub Codespaces GITHUB_TOKEN Leak
Orca Security researchers discovered RoguePilot, a flaw in GitHub Codespaces allowing attackers to steal GITHUB_TOKENs through indirect prompt injection.
Logic Flaws and Data Exfiltration in Autonomous AI Agent Architectures
Technical analysis of guardrail bypasses in LLM-integrated agents, highlighting the transition from conversational models to autonomous actors with privileged access.