Cisco announced at Cisco Live Melbourne that it is adding Anthropic’s restricted Claude Mythos model into Cisco AI Defense. Starting in preview for select buyers, the move expands Cisco’s work as a launch partner in Anthropic’s Project Glasswing defense push.
Because Claude Mythos is built to find deep software flaws, buyers won’t get an open chat prompt. Mythos runs locked inside Cisco’s test suite against owned, registered targets with named staff logins and hard stop switches.
Executive Summary
- Adaptive Probing: Claude Mythos drives multi-step attack plans inside Cisco AI Defense based on how target agents reply.
- No Open Prompts: Buyers never get raw access to query Mythos directly, cutting misuse risk.
- Owned Targets Only: Scans run strictly on pre-registered systems the buyer owns, blocking external web scans.
- Full Audit Trail: Named operator logins, compliance logs, and instant time limits govern every test run.
- Glasswing Roots: Extends Anthropic’s Project Glasswing into live business agents wired to internal APIs.
Moving From Static Prompt Lists to Multi-Step Agent Tests
Basic test suites fire canned trick prompts at chat windows to catch bad words. That checks little once firms hook AI agents directly into sales databases, billing APIs, and code repos.
Cisco AI Software chief DJ Sampath said teams must probe specific business uses, not generic chats. Using plain test rules, Mythos plans smart multi-step attacks that shift based on how an app reacts.
“You cannot scale what you cannot secure, govern, or observe.”
That runtime shift tracks alongside tools we covered in CrowdStrike’s AI kill switch and automated agent attack studies.
Strict Guardrails Around a Potent Bug-Finding Model
Handing every buyer an open bug-hunting model would invite abuse fast. Cisco SVP Jeff Schultz said AI Defense keeps Mythos locked behind strict guardrails instead of exposing raw prompts.
Staff must sign in by name, register owned endpoints first, and run timed tests that security chiefs can kill at once. Anthropic cyber lead Michael Moore noted firms want deep dry runs before putting agents in live work.
Banks are tightening developer gates in step, as seen when Goldman Sachs banned non-engineers from Claude Code.
What Security Chiefs Should Test During Preview
CISOs testing Cisco’s preview should probe tool calls, not just text blocks. Real risk hides when a rogue email tricks a helper agent into reading private keys or firing bad API hooks.
Clean operator logs and tight endpoint allowlists also give auditors clear proof before regulators check AI controls.
Security Controls: Cisco AI Defense + Claude Mythos
Core safeguards keeping Claude Mythos contained inside Cisco AI Defense:
| Control Area | Cisco Safeguard | Risk Blocked |
|---|---|---|
| Prompt Access | Hidden behind automated test workflows | Stops buyers using raw Mythos for external exploits |
| Target Scope | Pre-registered buyer systems only | Blocks scans against unowned websites |
| User Tracking | Named operator login & full logs | Keeps clear audit trail for compliance teams |
| Time Controls | Time-boxed runs with instant kill button | Halts risky test loops in staging apps |
| Test Depth | Plain-language adaptive attack planning | Finds chained tool bugs missed by canned lists |
Frequently Asked Questions
Why did Cisco add Claude Mythos to Cisco AI Defense?
Cisco added Mythos to run smart, multi-step attack tests against business AI agents before they launch live.
Can Cisco buyers prompt Claude Mythos directly?
No. Cisco blocks direct prompt access and confines Mythos strictly inside its red-teaming workflow.
How does Cisco stop Mythos from scanning random websites?
Tests only run on pre-approved systems owned by the customer, logged under named staff accounts with time limits.
When does the Claude Mythos preview launch?
Cisco unveiled the integration at Cisco Live Melbourne as a preview for select business clients.
[…] Hackers now blend open agent tools into live web scans, matching warnings on Linux edge server attacks. Defense vendors have rushed to match that pace, such as Cisco adding Anthropic red-teaming. […]