SecurityFirst 24 Hours: Responding to an AI Agent Incident
What to do when an autonomous agent acts beyond its bounds, hour by hour.
> Search
14 results for “prompt injection”
Security
SecurityMemory Poisoning: The New AI Attack Frontier
InjecMEM lets attackers plant persistent instructions in AI agents' memory with a single prompt.
SecurityGrok tricked into leaking user data
A researcher found that encrypting malicious instructions lets Grok exfiltrate user chats and personal details.
SecurityMCP Servers: A New Secret-Leak Vector
Model Context Protocol servers can expose enterprise secrets via plaintext configs, over-permissioning, and prompt injection, often undetected.
AI & MLOpenAI's Computer History logs your keystrokes
OpenAI's Computer History feature records clicks and typing, storing them unencrypted and expanding prompt injection risks.
AI & MLHow MCP Splits Can Leak Secrets
Malicious MCP servers can split instructions to make AI coding agents exfiltrate secrets, ASSET reports.
SecurityAtlassian Rovo Flaw Enabled Data Theft via Prompt Injection
A one-click prompt injection attack on Atlassian's Rovo could expose enterprise data across connected apps.
SecurityAI Agent Frameworks Face Security Crisis
Researchers reveal that vulnerabilities in AI agent foundations allow prompt injection to bypass critical trust boundaries.
SecurityAI Agents Deceived by Data Injection
Researchers identify a vulnerability where agents mistake untrusted input for verified facts, bypassing current security defenses.
SecurityAutomating the Hunt for Model Flaws
OpenAI is deploying an automated red-teaming model to aggressively stress-test its systems against persistent prompt injection risks.
SecurityCrowdStrike Exposes New Vulnerabilities in Generative AI Models
Researchers have identified five unique prompt injection techniques that exploit how large language models process user inputs.
SecurityThe Silent Sabotage: How MemGhost Alters AI Agent Memories
A new automated attack technique can plant persistent, invisible false memories into AI agents through a single, malicious email.
SecurityDefenders Weaponize Prompt Injection to Neutralize Rogue AI Agents
Researchers at Tracebit are utilizing context bombing to force AI models to trigger their own safety guardrails during attacks.
SecurityThe Image Blind Spot Letting AI Agents Steal Corporate Secrets
Researchers have demonstrated how bad actors can exploit AI code assistants by hiding malicious instructions inside unchecked images.