Cisco rolls out AI red teaming to hack your agents before attackers do

Cisco has added algorithmic red teaming and agent observability to Cisco AI Defense, its security platform for AI models and applications. Instead of leaning only on human testers, automated multi-agent systems probe models and agents for weaknesses, adjusting tactics as they go. The scans cover more than 200 risk categories, including prompt injection and data exfiltration.

The company frames the work as a full lifecycle play: discover the AI assets in use, detect threats against them, then defend by blocking those threats. Testing is tied to Cisco's Integrated AI Security & Safety Framework, which aligns with standards from NIST and MITRE. The platform also pulls real-time threat intelligence from Cisco Talos, Cisco's security research group, so tests draw on current attack techniques rather than a static checklist.

Cisco AI Defense first launched in January 2025. The February 10, 2026 update was described as the biggest expansion since the platform debuted, bringing the algorithmic red teaming to both agents and models. On March 23, 2026, Cisco released a free, self-service Explorer Edition that lets organizations evaluate AI models and agents in as little as 20 minutes with no upfront cost.

In June 2026, Cisco pushed further, adding Agent Validation and Agent Harness Testing to the Explorer Edition. Those features target agent-specific weaknesses, including tool routes and content channels. The Explorer Edition supports major agentic frameworks and systems that connect agents to outside services and data.

The focus on agents reflects a shift in what enterprises are deploying. Chatbots that answer questions are one thing; agents with access to tools and data can send files, call tools or act inside business systems. Cisco's newer tests probe indirect prompt injection, where malicious text hidden in emails, documents or web pages manipulates an agent reading that content.

Cisco has pitched the strategy at industry events, positioning it as a way to address risks from frontier AI while helping enterprises trust their own deployments.