Two announcements on November 9 describe the same problem from opposite ends. Cisco said at Cisco Live APJC in Melbourne that it will add Anthropic’s Claude Mythos to Cisco AI Defense, to attack customers’ AI applications and agents before someone else does. The same day, the Linux Foundation published research, sponsored by ActiveState, showing that most security teams can no longer keep up with the code their developers ship with AI help.
One is a vendor selling a stronger attacker to defenders. The other is a survey of defenders who say they are already outpaced.
What Cisco is adding
Cisco AI Defense launched in January 2025, and Cisco says it has helped hundreds of customers secure AI applications since. The new piece is adaptive red teaming: Mythos will probe a customer’s AI apps and agents, adjust its attacks based on how they respond, and report what broke.
Cisco is careful about the guardrails, which is a sign of how capable it thinks the model is. Testing runs only against targets the customer has registered and is authorized to test. There is no open-ended scanning of the internet. Each session has a named, authenticated operator, objectives set by a human, a fixed time window, and can be stopped at any point. Availability is “coming soon” for selected enterprises. Cisco gave no date and no pricing.
The rest of the announcement is about watching agents in production through Splunk:
- Tokenomics, which tracks what AI agents cost to run, goes to private preview in December 2026 and general availability on AWS Marketplace in early 2027.
- Eval-G1, a set of small evaluation models, is generally available this month with an Evaluation Tuner. Cisco says they cost up to 98% less than using a commercial LLM as the judge and run 2.5 times faster.
- Cisco IQ, which covers about 17,000 customers and more than 22 million devices, now feeds Splunk, with MCP and ServiceNow connectors.
What the Linux Foundation found
The research surveyed 172 security practitioners, with OpenSSF and LF AI & Data. The headline is that 58% say code now ships faster than their team can review it. Sixty percent say AI coding assistants have increased the number of outside packages in their code, and transitive dependencies, the packages those packages pull in, get reviewed only about half the time.
AI helps the defenders too. Respondents say AI tools found 30% of the vulnerabilities they caught last year, 69% say their security posture improved, and work went 17% faster. But false positives added about 30% to the workload, and 82% worry that leaning on AI is eroding developers’ own security skills.
The top concerns were supply chain attacks (93%), prompt injection (90%) and the sheer volume of updates (87%). The widest gap between worry and readiness, 41 points, was on AI-automated exploit generation and fuzzing: the exact capability Cisco is now packaging for defenders.
Organization matters. Only 30% have someone whose job is AI security, and about one in three have an AI security policy. Of teams with a dedicated role, 75% feel prepared, against 50% overall. Even among teams with a policy, only 57% require a human to review AI-generated code. The report notes that this is correlation, not proof that hiring the role causes the readiness.
What buyers should take from it
The survey says security teams are short on review capacity and most afraid of automated attacks. Cisco’s pitch is to fight automation with automation, under tight controls. That is a reasonable answer to the second problem but not the first. A red-team agent finds weaknesses; someone still has to triage and fix them, and the survey shows that is where teams are already stretched.
For a CISO, the cheaper lessons from the research come first: name an owner for AI security, write the policy, and require human review of AI-written code. Those steps don’t need a model. Adaptive red teaming makes sense once there is a team ready to act on what it finds. With no date or price from Cisco, there is time to get that in place.