Mythos Inside the Wrapper: What Cisco's AI Defense Deal Says About Shipping an Offensive Model

Since yesterday the headlines have said “Cisco adds Anthropic’s Mythos to its security products”. That is accurate, but it is not the interesting part. The interesting part is how Mythos is being delivered: customers never get the model. They get a tightly fenced capability inside a vendor’s product, and the fence is where the engineering is.

What was announced

At Cisco Live APJC in Melbourne on 9 November, Cisco said it is adding Claude Mythos to AI Defense, its product for testing AI models, applications and agents before they reach production. AI Defense already has an adaptive red teaming framework that builds multi-step attacks against a target system. Mythos is described as an addition to that framework: it plans the tests and turns the results into attack ideas tailored to each system.

Availability is vague. Cisco says adaptive red teaming with Mythos is coming soon for select enterprises, with no date. The Register reports that Cisco will offer it under existing AI Defense licences at no extra cost, which comes from a Cisco executive in that interview rather than the press release, so treat it as reported, not contractual.

The same announcement included Splunk Agent Observability updates (a cost feature called Tokenomics, in private preview from December) and Cisco IQ changes. Those are worth a read if you run Splunk, but they are a separate story.

The fence

Cisco lists four safeguards. Reading them as a network engineer, they look like a change-control process applied to an offensive tool:

SafeguardWhat it means in practice
Registered, authorised targetsIt only works against systems the customer has registered, owns and authorised for testing. It cannot scan arbitrary internet infrastructure.
Named operator attributionEvery engagement requires authentication and is cryptographically tied to a named person at the customer.
Revocable executionEach engagement has a fixed job and objectives set by a human, runs in a time-bound session, and can be halted on demand.
Dual-layer oversightProduct-level controls in AI Defense sit alongside program-level safeguards and monitoring.

On top of that, customers have no open-ended prompt access. They use a structured interface, every testing action is recorded for audit, and results can be carried over into follow-up attack loops.

Take the same four ideas and drop them into a firewall change process and none of it would look out of place: scoped targets, an accountable owner, a time window and a kill switch. What is new is applying them to a model that can plan attacks.

Why the delivery model matters

Mythos is not generally available. Anthropic first released it through Project Glasswing to vetted organisations, and Cisco joined that programme in April. The Register’s reading is that Cisco and IBM now expose Mythos’s functionality to end users without letting them use the model itself, and that this suits the risk: in the wrong hands, a model that is good at finding and exploiting software flaws helps attackers too.

The tight control is not theoretical. The Next Web notes that in April an unauthorised group reached Mythos through a contractor’s access. I have not verified the details of that incident beyond that report, but it is a fair illustration of why a vendor wrapper with named operators and audit trails is more than box-ticking. The weakest point in a controlled-access scheme is usually a person or a credential, not the model.

It is not just Cisco

IBM announced the same day that it plans to bring Mythos into Concert Protect, with the Next Web reporting early 2027. IBM says it will use it to find weaknesses traditional scanners miss, work out which ones pose real risk, and generate fixes for developer approval. The Next Web also notes that Palo Alto Networks launched always-on testing built on Mythos and an OpenAI cyber model in September. Anthropic, asked by The Register whether it has similar deals with other third parties, declined to say.

So the pattern is a distribution channel: a model Anthropic will not hand out directly, reaching enterprises through products they already pay for. That is a commercial decision as much as a safety one, and both readings can be true.

What is not public

It is worth being clear about what the announcements do not tell you:

  • No date and no customer list. “Select enterprises” and “coming soon” are the whole timeline.
  • No performance numbers. I have seen no public benchmark showing how much better Mythos-driven red teaming is than what AI Defense already does. Cisco researchers did use Mythos in July to push 15 chatbots past their guardrails, according to the Next Web, but that is a research result, not a product measurement.
  • No detail on the program-level safeguards. The product-level controls are described. What the “program-level” layer actually checks is not.
  • Scope is AI systems. This is red teaming for AI apps and agents, not a general vulnerability scanner for your network estate. IBM’s Concert Protect is the closer fit for application scanning.

What I would ask a vendor

If a security vendor offers you a frontier model inside a managed wrapper, these are the questions I would want answered before relying on it:

  1. What does “registered target” mean technically? How is ownership proved, and what stops a typo or a stale registration pointing it at someone else’s system?
  2. Where do the audit logs live, and can I export them? A log you cannot get out of the vendor is not much use in an incident review.
  3. What happens to my findings? Test results about how to break my agents are sensitive. Who at the vendor and the model provider can see them, and for how long?
  4. How do I stop a run, and how fast does it stop? “Can be halted on demand” is a claim to test.
  5. What does it do that the existing red teaming did not? Ask for a before-and-after on a system you know well.

For the wider context on testing AI agents and what “production-safe” really means, see AI Pentest Agents Part 5, and AI Part 4 covers the rails I put around my own Claude access to this site. The Judgment Gap Just Got Numbers is useful for calibrating how much to trust any AI-driven testing claim.

My take

The model is the headline, but the wrapper is the product. If frontier offensive-security models are going to reach enterprises at all, I would rather they arrive like this, scoped, attributable and revocable, than as an API key. Whether the safeguards hold up is something only independent testing will show, and nothing announced so far lets us do that yet.

A note on disclosure

I use Claude to help run this site, as I wrote in How I Use Claude to Help Run This Blog, and Claude helped research and draft this post. This is a story about Anthropic’s model, so weigh my take accordingly. The facts above come from the primary announcements and the trade press linked below, and where I could not verify something I have said so.

Sources: Cisco press release, 9 November 2026, The Register, The Next Web, Cisco: Rising to the Era of AI-Powered Cyber Defense, StreetInsider.