On October 5, Zscaler announced a limited preview of Anthropic's Claude Mythos 5.1 inside Zscaler Endpoint AI Security. This one came out as a blog post from Dhawal Sharma, Executive Vice President of AI Security and Strategic Initiatives, not a press release, which is how it slipped past me for a week. Standard disclosure: I work at Zscaler, so read the enthusiasm as informed, not neutral.
Shadow AI Moved Onto the Laptop
The problem framing in the post is one every security team will recognize. AI adoption is happening bottom-up. Employees grab the latest assistant, coding agent, AI browser, or dev tool because it helps them get work done, and they do it long before anyone in security has evaluated it.
"The result is a sprawl of shadow AI: a fast-growing layer of applications, agents, extensions, models, packages, and configuration files on employee devices."
— Dhawal Sharma, Executive Vice President, AI Security and Strategic Initiatives, Zscaler
Here's what makes agents different from the shadow IT we've been fighting for a decade: they act. An agent can use your identity and permissions, call tools, read files, and run commands on your behalf. And because it's acting as you, the audit log often just says it was you. Try running an investigation when you can't tell what the person did from what their agent did.
Traditional endpoint tools weren't built for this. MDM knows what's installed. EDR knows what processes ran. Neither has much context about which agent acted, what task it was doing, which tools it called, and what data it touched.
What Endpoint AI Security Does
Zscaler's answer has three parts, per the post:
- Discovery: find the AI assistants, agents, browsers, dev tools, models, packages, extensions, and connected services across managed devices.
- Runtime visibility: capture what agents actually do, including prompts, tool calls, file access, commands, and network activity.
- Policy enforcement: restrict unauthorized agents, block dangerous commands, and protect sensitive files, right at the endpoint.
Where Mythos Comes In
The Mythos 5.1 piece powers a capability called Exploit Paths. The post describes it as cross-correlating evidence to identify potential threats, explain how the steps connect, and show defenders where to interrupt the sequence.
That's the right job for a frontier model. Individual endpoint events are usually boring on their own. An agent with broad permissions? Common. An agent calling a third-party tool? Normal. An agent touching a sensitive file? Happens all day. But all three together, in sequence, is an attack path. Humans are bad at spotting that kind of chain in a pile of separate alerts. A model that can reason over the whole sequence is not.
The part I appreciate most is the honesty built into the output. Exploit Paths shows the supporting evidence, confidence levels, which steps are inferred versus confirmed, and the defensive intervention points. That's the difference between an AI security tool you can act on and one that just hands you a scary-sounding verdict.
Bottom Line
AI agents are becoming coworkers with admin rights and no supervisor. Zscaler using Anthropic's most capable model to connect the dots between what those agents do on a laptop is exactly the kind of "use AI to secure AI" move I want to see, as long as it shows its work, and this does. It's in limited preview now, so this is an early look, not a GA launch.
Full post: Bringing Anthropic's Claude Mythos 5.1 to Zscaler Endpoint AI Security