Cisco AI Defense Integrates Anthropic Claude Mythos for Advanced AI Red Teaming

Cisco Brings Frontier AI Red Teaming to Enterprise Security

Cisco announced at Cisco Live in Melbourne that it's integrating Anthropic's Claude Mythos into its AI Defense platform, becoming one of the first AI security products to offer the bug-finding AI model's capabilities for adaptive AI red teaming3

3

. The integration allows enterprises to test AI agents and applications for security vulnerabilities before production deployment, addressing the growing need to secure agentic AI systems that connect to complex networks of tools and sensitive resources.

"You need to really deeply understand each of [the] use cases so that you're red teaming against those use cases to be able to discover those gaps," DJ Sampath, Cisco's senior vice president and general manager of AI Software and Platform group, told CRN4

4

. "We're moving from generic red teaming to adaptive red teaming, but more importantly, we're using Mythos model to be able to do it."

How Adaptive AI Red Teaming Works with Claude Mythos

Cisco AI Defense already offers two types of testing: standard red teaming that assesses common vulnerabilities like prompt injection, jailbreaks, and sensitive data exposure across thousands of categories, and adaptive red teaming where users define custom testing criteria in plain language3

3

. The system analyzes target applications, develops tailored attack scenarios, and refines its approach based on responses.

Claude Mythos augments this adaptive AI red teaming engine with frontier-level attacker reasoning. "By bringing Claude Mythos to AI Defense, we're augmenting that system to deliver an advanced AI red teaming experience to customers," Cisco stated3

3

. The model generates sophisticated attack ideas and multi-step attack paths while operating under strict safeguards designed for authorized testing AI agents.

Controlled Access Without Direct Model Exposure

Customers cannot prompt Mythos directly. Instead, the model sits inside AI Defense with a fixed job, registered targets, and objectives defined by human users2

2

. Every test is tied to a named operator and fully auditable, ensuring adversarial testing mirrors real threats without compromising customer environments3

3

.

"Customers get no open prompt access, every test is logged to a named person," according to The Next Web2

2

. Mythos cannot scan the wider internet and works only on systems the customer owns and has registered for testing. This controlled approach matters because unrestricted access to the bug-finding AI model could help attackers exploit software flaws.

Natalie Mead, Anthropic's head of applied AI for Asia-Pacific, said the company has a long queue of buyers wanting Mythos access, and working with Cisco helps satisfy that demand1

1

. The integration comes at no additional cost under existing AI Defense licenses1

1

.

IBM Joins the Mythos Integration Wave

Hours after Cisco's announcement, IBM revealed it will integrate Claude Mythos into its Concert tool, which uses AI to scan applications for risk and security vulnerabilities1

1

. IBM plans to use Mythos "to uncover weaknesses traditional scanners miss, determine which pose real risk and generate fixes for developer approval," with availability expected in early 20272

2

.

This follows Palo Alto Networks launching always-on testing on Mythos and OpenAI's cyber model in September2

2

. The pattern suggests Anthropic is expanding Mythos access through partner integrations rather than direct distribution, maintaining tight controls while meeting enterprise demand.

Enhanced Performance and Security in Mythos 5.2

Anthropic released version 5.2 of Mythos alongside the Cisco announcement, delivering outputs up to 30 percent faster than version 5.1 at reduced costs of $8 per million input tokens and $40 per million output tokens1

1

. Users can activate "fast mode" for token generation up to 2.5 times faster at doubled prices of $16 per million input tokens and $80 per million output tokens.

Anthropic conducted extensive prerelease safety testing, including novel evaluations for concerning behaviors like sandbox escapes. The company concluded "risks from misalignment remain low" and implemented "strongest-ever protections against distillation attacks"1

1

. This addresses concerns about Chinese AI labs potentially using distillation to clone Anthropic's models without expensive training.

What This Means for Enterprise AI Security

The integration addresses a critical gap as businesses rapidly adopt agentic technologies. "Enterprises are increasingly building AI agents into their core operations and want to test those agents thoroughly before putting them into production," said Michael Moore, Anthropic's cybersecurity lead2

2

.

Cisco AI Defense, launched in January 2025, now serves approximately 400 to 500 customers4

4

. The adaptive red teaming with Claude Mythos will roll out first to select enterprise customers, with broader availability planned but no firm dates announced4

4

.

For partners, Sampath called it "a game changer," noting that no other red teaming service offers a model this powerful for testing AI agents across client environments4

4

. As AI agents move beyond cloud environments to laptops and edge devices, Cisco plans to expand AI Defense capabilities to secure deployments across all these environments.