Prompt Injection

Anthropic says Claude Opus 5 and Auto Mode reduced browser prompt injection success to zero in internal tests, a notable claim for AI agents.

OpenAI introduced GPT-Red, an automated red-teaming system it says improves prompt injection defenses and could help safety scale with model capability.

Researchers exploited role-model prompt injection to bypass LLM safety guardrails and extract harmful content including drug synthesis instructions.

Prompt injection, ranked OWASP LLM01, is evolving to target multi-agent systems, RAG supply chains, and model routers in enterprise AI deployments.

Threat actors exploited legitimate AI security tools at 90+ organizations using malicious prompt injection, with the next wave gaining write access to firewalls.

Oasis Security researchers discovered three chained flaws in Anthropic's Claude — including a prompt injection, Files API exfiltration path, and open redirect — enabling silent data theft through a Google Search ad.

Security researchers demonstrated that an autonomous AI agent successfully compromised McKinsey's internal AI system in less than two hours by exploiting prompt injection—a well-known but still widely unmitigated attack vector—raising urgent concerns about enterprise AI security.

Cybersecurity experts warn Moltbook, a social network for AI agents, poses prompt injection risks that could compromise thousands of agents simultaneously.

New enterprise security features protect against AI prompt injection and data exfiltration with deterministic controls for high-risk users.