#prompt injection
77 stories taggedprompt injection · page 5 of 6.
Estonia Wants to Give AI Agents Government-Issued IDs — With Spelled-Out Permission Scopes
The Baltic nation's AI Council is proposing state-backed digital identities for AI agents, defining exactly what they're allowed to do before they touch your data or your bank account.
SearchLeak: How a microsoft.com Link Could Have Drained a Copilot Tenant
Varonis Threat Labs chained three bugs in Microsoft 365 Copilot Enterprise Search into a one-click exfil path that lived behind a trusted Microsoft URL.
Agentjacking: Poisoned Sentry Error Reports Hijack AI Coding Assistants
Researchers describe a prompt-injection class that turns developer error-tracking pipelines into a remote code execution path against AI coding agents.
AI Web Agents Have No Reliable Prompt Injection Defenses, Benchmark Finds
Researchers ran 3,168 adversarial tests against GPT-5 and Gemini-powered agents. The 'Robust Behavior' outcome — agent completes task, attacker gets nothing — never appeared.
Researchers Turn OpenClaw Into a Confused Deputy With Hidden Prompts
Two teams show the self-hosted AI agent will execute attacker instructions smuggled inside contacts, location pins, and other benign-looking inputs.
Twelve Controls That Actually Matter Once AI Ships to Production
Visibility into AI applications is a starting point, not a security posture. Here is what ongoing monitoring and defense of production AI systems looks like in practice.
AI Red Teaming Grew Up. The Job Description Is Still Being Written.
The tools broke when LLMs arrived. Now the discipline is rebuilding itself in real time — and the threat model includes teenagers with too much free time.
OpenAI's Lockdown Mode Admits the Problem It Can't Quite Fix
The new containment feature reduces AI-enabled data exfiltration — it doesn't stop it. Experts are divided on whether enterprises should even trust a vendor to police itself.
OpenAI Ships ChatGPT 'Lockdown Mode' to Blunt Prompt-Injection Data Theft
The opt-in setting strips connectors and browsing tools that attackers have used to siphon data from logged-in sessions.
One GitHub Issue Was Enough to Pwn Repos Running Claude Code Action
A bug in Anthropic's Claude Code GitHub Action turned issue triage into arbitrary code execution — including, briefly, against the action's own repo.
A Single Notification Could Hijack Gemini on Android
Researchers showed how a poisoned WhatsApp, Slack or SMS alert could weaponize Google's voice assistant — no malicious app required.
Someone Finally Tested 100 AI Agents for Security. Here's the Framework They Used.
A new evaluation methodology ranks AI agents by vulnerability, blast radius, and defensive posture. The results are a useful corrective to vendor claims.
ChatGPhish: When ChatGPT's Markdown Renderer Becomes a Phishing Vector
Permiso researchers show how implicit trust in Markdown links and images inside ChatGPT responses turns the assistant into a credible delivery surface for prompt injection and credential theft.
You Can't Audit What You Can't See: The Agent Governance Hole Nobody Wants to Talk About
Enterprises are shipping AI agents into production without inventories, without trace pipelines, and without a coherent answer to a basic question: what is this thing actually doing?
Anthropic Wires Claude Into 28 Enterprise Security Platforms
The AI company is pushing deeper into corporate security stacks, connecting Claude to vendors including CrowdStrike, Okta, and Zscaler.