The cybersecurity industry has a dirty secret: the defenders are always one step behind. Attackers only need to find one way in. Defenders need to find every way in, before anyone else does. It is an exhausting, asymmetric war of attrition, fought with too few people and too many alerts. In 2026, Cisco is betting that the only way to close that gap is to hand the fight to a machine that thinks like an attacker—and it is turning to Anthropic to build it.
On November 9, 2026, at Cisco Live in Melbourne, Cisco announced that it is integrating Anthropic’s Claude Mythos—the security-focused frontier model that Anthropic has deliberately restricted from public access—into its AI Defense platform for adaptive AI red teaming . It is one of the first AI security products to use a model of this caliber for the specific purpose of attacking AI systems to find their weaknesses before malicious actors do.
The move raises a provocative question: can AI become the ultimate digital bodyguard? The answer, as with most things in cybersecurity, is more complicated than a simple yes.
What Cisco Is Actually Doing
Cisco AI Defense is not a new product. Since its launch, it has combined asset visibility, supply chain scanning, algorithmic red teaming, and runtime guardrails into a single platform. What is new is the engine powering the adaptive red teaming capability.
Red teaming is the practice of simulating attacks on a system to identify vulnerabilities before they can be exploited. Cisco’s AI Defense already offers two types: standard red teaming, which tests for common vulnerabilities like prompt injection and jailbreaks across thousands of categories, and adaptive red teaming, which allows customers to define custom testing criteria in plain language and then uses an agentic framework to analyze the target, develop tailored attacks, and refine its approach based on the system’s responses.
Claude Mythos is now being added to the adaptive red teaming engine. Its role is specific: to plan tests and generate attack ideas for each system. As Cisco’s Amy Chang and Vikram Varakantam put it, “Bringing Claude Mythos to AI Defense augments the adaptive red teaming engine where the attacker’s reasoning matters most. As the engine works to understand a target system and plan tests, the model’s reasoning produces attack ideas. When those ideas become concrete payloads, the benefit is sharper still”.
In other words, Mythos is not replacing the red teaming framework. It is providing the adversarial creativity that makes the framework smarter—the ability to think like a hacker, to spot multi-step attack paths that conventional testing might miss, and to reason through how a specific AI agent, with its specific configuration and access to specific resources, might be manipulated into leaking data, misusing tools, or bypassing its guardrails.
Why Mythos Matters: The Model That Was Too Dangerous to Release
Claude Mythos is not a general-purpose chatbot. It is a security-specialised frontier model that Anthropic has deliberately kept out of public hands because of its offensive capabilities. When Anthropic released Mythos 5.2 in November 2026, it was made available only to vetted organizations through Anthropic’s Cyber and Life Sciences Verification Programs—not to the general public, and not through an open API.
The reasoning is straightforward. Mythos is exceptionally good at finding vulnerabilities. In testing, it has been used to reproduce 81 previously unknown high-severity vulnerabilities in Firefox and to develop working exploits for 948 Linux kernel vulnerabilities. In Cisco’s own research, the model was used to push 15 different chatbots past their guardrails in a matter of turns.
That capability is precisely what makes it valuable for defense. A defender who can think like an attacker—who can generate the same attack ideas, explore the same multi-step paths, and identify the same weaknesses—is a defender who can fix the problems before they are exploited. As Michael Moore, Anthropic’s cybersecurity lead, put it: “Enterprises are increasingly building AI agents into their core operations and want to test those agents thoroughly before putting them into production. By bringing Claude Mythos to Cisco AI Defense, Cisco can help security teams red team their agents with advanced reasoning that plans tests and generates attack ideas for each system”.
The Safeguards: A Bodyguard With a Strict Rulebook
The obvious concern is that a model capable of generating attacks could be misused. Cisco and Anthropic have built a tightly controlled deployment model to address that risk.
Customers cannot prompt Mythos directly. There is no open interface, no chat box, no API that a customer can query freely. Instead, Mythos operates within AI Defense’s structured framework, with a fixed job and human-defined objectives.
Testing is restricted to authorized targets. Mythos can only operate against systems that the customer has registered and is explicitly authorized to test. It cannot scan the wider internet or attack unregistered infrastructure.
Every engagement is tied to a named operator. Tests are authenticated, linked to a specific person, and recorded for audit. The activity can be stopped at any time.
It is initially available only to select enterprises. Cisco is rolling out the integration gradually, with no date set for general availability.
Cisco frames this as “frontier attacker reasoning with responsible controls”—a deliberate effort to make the most powerful security-relevant AI capabilities available to defenders under verified, controlled access rather than withholding them entirely. It is a bet that the risk of giving defenders a powerful tool is outweighed by the risk of leaving them without one.
The Broader Shift: AI vs. AI
Cisco’s move is not happening in isolation. It is part of a broader recognition across the security industry that the threat landscape has fundamentally changed.
AI-powered cyberattacks are no longer a future concern. Deepfake-enabled financial fraud has grown into a $40 billion annual market. AI agents have been used to find and exploit vulnerabilities at machine speed. The “zero-day clock”—the time between a vulnerability’s disclosure and the first attack—has collapsed from an average of 771 days in 2018 to mere hours today. Traditional defenses, built around human analysts reviewing alerts and patching systems on a weekly cycle, cannot keep pace.
The response from the security industry has been a wave of AI-powered defensive tools. Palo Alto Networks has launched always-on testing powered by Mythos and OpenAI’s cyber model. IBM plans to add Mythos to its Concert Protect tool in early 2027. Cisco itself is expanding its AI security portfolio beyond red teaming, adding Splunk Agent Observability to monitor AI agent performance and cost, and new infrastructure security capabilities through Cisco IQ and Cisco Data Factory.
The common thread is a recognition that AI is both the problem and the solution. The same technology that allows attackers to automate vulnerability discovery and craft convincing deepfakes can be used to automate defense, to simulate attacks, and to find weaknesses before they can be exploited.
The Limits: A Bodyguard, Not a Silver Bullet
For all its promise, Claude Mythos in Cisco AI Defense is not a complete solution. It is a powerful tool with a narrow scope.
It tests AI systems, not everything. The focus is on AI applications and agents—the systems that are increasingly being deployed with access to sensitive data and business workflows. It does not replace traditional security testing for networks, endpoints, or cloud infrastructure.
It finds vulnerabilities; it does not fix them. The output of a red teaming engagement is a report of findings and recommended follow-up actions. The actual remediation—the patching, the reconfiguration, the policy changes—still requires human judgment and effort.
It is only as good as the targets it is given. If a customer has not registered a system for testing, Mythos cannot find its vulnerabilities. The scope of protection is defined by the scope of authorization.
It is not a substitute for security hygiene. No amount of AI-powered red teaming can compensate for weak passwords, unpatched software, or a lack of basic security controls. The bodyguard can find the unlocked windows, but someone still has to lock them.
A Bodyguard That Never Sleeps
Can AI become the ultimate digital bodyguard? The honest answer is that it is already becoming one—but not in the way the phrase suggests.
It is not an invincible shield that stops every attack. It is a tireless, intelligent, and relentlessly adversarial sparring partner that finds the gaps in your defenses before someone else does. It is a tool that compresses the time between vulnerability and patch, that thinks like an attacker so that defenders can think one step ahead, and that operates at the speed and scale of the threats it is designed to counter.
Cisco’s integration of Claude Mythos into AI Defense is a significant step in that direction. It brings a frontier security model into a controlled, auditable, and authorized environment, and it gives enterprises a way to test their AI agents against the kind of adversarial reasoning that only a model of Mythos’s caliber can provide.
The ultimate digital bodyguard is not a single product or a single model. It is a layered system of defenses—AI-powered red teaming, runtime guardrails, observability, and human expertise—working together to protect an attack surface that is growing more complex by the day. Cisco and Anthropic have built a powerful new layer. Whether it becomes the foundation of a new era of AI-driven defense depends on how well the industry learns to use it.
The attackers are already using AI. The defenders finally have a reason to fight back.