Anthropic dominates Cisco’s new AI security rankings

Giving an AI agent access to your corporate database is like handing your company keys to a brilliant intern who might spill company secrets after five minutes of smooth talking.

To see which models actually hold their ground under pressure, Cisco just published its free LLM Security Leaderboard. The outcome was a total blowout by Anthropic.

Anthropic swept eight of the top ten positions. Claude Opus 4.5 clinched the top spot, followed closely by Sonnet 4.5 and Haiku 4.5. OpenAI managed just two spots in the top ten, with GPT-5.2 landing in seventh and GPT 5 Nano placing ninth.

The bottom of the table was rough for open and smaller architectures. Mistral anchored the absolute bottom with Magistral Small 2509 and Ministral 3 14b Instruct. The bottom ten also swept up models from DeepSeek, Cohere, Qwen, and xAI.

Real attacks don't happen in a single prompt

Most safety benchmarks only test single-turn jailbreaks, basically tossing one bad prompt at a model and seeing if it flinches. Cisco went after multi-turn conversational attacks instead.

Amy Chang, Cisco's Head of AI Threat Intelligence & Security Research, pointed out that real attackers do not give up after one try. They wear down the model guardrails over long, deceptive conversations.

Cisco also kept things raw. No extra safety filters or third-party guardrails were applied to rescue any model during the benchmark, forcing each LLM to rely strictly on its native alignment.

According to Cisco's AI Readiness Index, 83% of organizations plan to deploy agentic AI, yet only 29% feel prepared to do so securely. Companies want agents to do real work, but the security gap is widening fast.

Why it matters

If you are planning to deploy AI agents that have read or write access to internal tools, this benchmark is a reality check.

Anthropic's massive lead proves that their obsession with Constitutional AI and native alignment wasn't just corporate virtue signaling. It translates into real resilience when long-context attacks try to trick the system. Choosing a cheaper or less-aligned model might save API costs today, but it could cost you dearly if an agent gets socially engineered over multi-turn interactions.

We also think Cisco's public approach puts real pressure on the broader ecosystem. While cloud providers like Microsoft have shared internal safety ratings for select enterprise customers, Cisco put this data out in the open for free. When enterprise buyers can clearly see who landed in the bottom ten, labs like Mistral, Cohere, and DeepSeek will be forced to treat adversarial robustness as a core product feature rather than an afterthought.