Claude Cowork Security Risks: What Enterprise Teams Need to Control

Built for Speed: ~10ms Latency, Even Under Load
Blazingly fast way to build, track and deploy your models!
- Handles 350+ RPS on just 1 vCPU — no tuning needed
- Production-ready with full enterprise support
Two days after Claude Cowork launched in January 2026, researchers at PromptArmor showed what an agent on a work laptop can be talked into. A Word document with 1-point white text carried hidden instructions. Cowork read the file during a normal task, found the user's financial documents, and uploaded them to an account the attacker controlled, using an API that the sandbox still allowed. The user did nothing except open a file.
That is the core issue. Cowork reads content it didn't write, and that content can contain instructions. The difference from a chatbot is what happens next: Cowork can reach local files, the user's logged-in browser, connectors like Slack and Google Workspace, and scheduled tasks that keep running after the user walks away.
This guide covers the six attack surfaces, what Anthropic now offers for audit and monitoring, and the controls to put in place before you roll Cowork out beyond a pilot. We've updated it for what changed in 2026, including Anthropic's Compliance API now covering Cowork.
What Claude Cowork Actually Is
Think of Cowork as an agent process working on behalf of each employee. It can run code in a sandbox, read and write files in folders you let it mount, browse the web, call MCP servers and native connectors, and run scheduled tasks. On some plans it can also use Computer Use to control the desktop directly.
A successful injection against a chat assistant leaks the conversation. A successful injection against Cowork can send files out, post messages as the user, create a task that repeats nightly, and touch every service the user is signed in to. Plan for that, not for a smarter chatbot.
Claude Cowork Security Risks at a Glance
The Six Attack Surfaces
1. Prompt injection through files and web pages
This is the most likely and most severe risk. A user asks for something ordinary. During the task Cowork reads a document, web page, email or API response that contains hidden instructions, and acts on them. Common hiding places are white or 1-point text in Office files and PDFs, hidden elements on web pages, email bodies, and records returned by database queries.
Anthropic has reported roughly a 1% attack success rate for Claude in Chrome against its own strongest internal attacker, even with mitigations in place. One percent sounds small until you multiply it by thousands of sessions a day. Treat it as a risk to contain, not one that settings remove.
2. The browser acts as the user
When Cowork browses through an authenticated browser context, it inherits whatever the user is signed in to: Google Workspace, Salesforce, internal tools, cloud consoles. A malicious page can tell it to read data from one of those sites or submit a form on the user's behalf. Turn web search off for people who don't need it, use a domain allowlist for those who do, and route traffic through your proxy so your CASB and DLP tools can see where it goes.
3. Scheduled tasks make an injection persistent
Dispatch lets Cowork run tasks while the user is away. That is useful, and it is also how a single injection becomes a standing problem: an injected instruction creates a nightly task that keeps reading files and sending them out. Keep Dispatch off by default, alert on every new scheduled task, and audit existing tasks when you assess a deployment.
4. Connectors and MCP servers widen the blast radius
Every connector and MCP server adds to what a hijacked session can reach. Native connectors inherit the user's full permissions, so a Slack connector that can post as the user is an exfiltration path. A sensible default is read-only access to Slack, Google Drive and Calendar, SharePoint and Teams, and GitHub, with write access scoped to named teams or repositories, and database MCP servers on a read-only role. Our guide to the best MCP servers for Claude Code covers which servers are worth the risk.
5. Computer Use skips the usual permission checks
Where it's available, Computer Use drives the mouse and keyboard directly, outside the sandbox. An injection that reaches it can touch any app the user has open, including a password manager or an SSH terminal. Keep it disabled for enterprise users until you can scope it per application.
6. You can't always see what the agent did
Visibility is where the story has changed most this year, so it gets its own section below.
Audit and Monitoring: What Has Changed
Earlier in 2026, Cowork activity sat outside Anthropic's audit logs, Compliance API and data exports, and many guides, including an earlier version of this one, called that the biggest gap. Anthropic has since closed much of it. In August 2026 it announced that the Compliance API covers Cowork on desktop, web and mobile for Claude Enterprise customers, including session content and verified user metadata. Anthropic marked that coverage generally available on August 26, 2026.
That doesn't make everything visible. Some gaps remain:
- Plan coverage. Compliance API coverage is for Enterprise customers. Check what your plan includes before you promise auditors anything.
- Some sessions are outside it. Anthropic lists Claude Code on the web, Claude Code through the Claude Platform, and sessions on Bedrock, Vertex AI and Microsoft Foundry as not covered.
- OpenTelemetry is metadata by default. Export is available on Team and Enterprise plans and is set up under Organization settings. It records tool names, approvals, model, token counts and cost, but prompt text and tool inputs stay off unless you set the content-capture device policy. If you turn content on, decide who can read it and how long you keep it, because it can contain sensitive data.
- History lives on the device. Conversation history is stored locally, so disk encryption and endpoint security become your data-at-rest controls.
The practical answer is to layer sources: the Compliance API for the audit trail, OpenTelemetry for operational alerts in your SIEM, endpoint and network logs, and tool-level logs from a gateway. Alert on the events that matter most: new scheduled tasks, connections to MCP servers that aren't on your allowlist, connector write actions at odd hours, reads from sensitive directories, and bypass-permission modes.
Lock Down Cowork Before You Launch
Identity
Enforce SSO through your identity provider so MFA comes with it, set up SCIM so offboarding removes access automatically, and use domain capture so corporate emails land in the enterprise workspace. On managed devices, add tenant restrictions by injecting the anthropic-allowed-org-ids header at your proxy, which stops people signing in to personal Claude accounts and sidestepping your controls.
Managed settings and egress
Push managed settings through your MDM so users can't override them. Disable bypass-permissions mode, deny reads of credential locations such as .env, .ssh and .aws, require confirmation for writes, and allow only approved MCP servers. In the admin console, configure network egress with a short allowlist. Anthropic added admin control over which hosts Cowork can fetch in May 2026, and it matters because the January attack worked through an API the sandbox still reached.
Pick a posture and write it down
PromptArmor's guidance for Cowork offers a useful way to frame it. Three postures cover most organizations:
Start in the posture that matches your most sensitive data, then loosen it for specific teams with a documented need.
Where a Gateway Fits
Anthropic's controls cover what you can configure in Cowork. Most companies also run Claude Code, Cursor, internal agents and several model providers, each with its own permission model and logs. A gateway adds one layer that works the same across all of them.
TrueFoundry's AI Gateway handles the model-call layer, with per-user attribution, budgets, rate limits, model access control and request-level guardrails. The MCP Gateway handles the tool layer: a registry of approved servers, role-based access to specific tools, checks before execution, and an audit trail of every call with user, tool, payload and response. Both export through OpenTelemetry to your SIEM, and they sit alongside the Compliance API, not in place of it. See our guides on enterprise security for Claude and governing Claude Code with an AI gateway for setup details, and check what is supported for your Cowork deployment mode.
Claude Cowork Security Checklist
- Inventory first. Find every Cowork deployment, including people on personal accounts, and the connectors and MCP servers attached.
- Lock identity. SSO, SCIM, domain capture and tenant restrictions.
- Deploy managed settings. Bypass off, credential paths denied, approved MCP servers only.
- Restrict the network. Egress allowlist, web search off where unneeded, traffic through your proxy.
- Least privilege on connectors. Read-only by default, write access by exception.
- Dispatch and Computer Use off until a team has a documented need.
- Turn on monitoring. Compliance API on Enterprise, OpenTelemetry to your SIEM, alerts on scheduled-task creation and sensitive reads.
- Protect devices. Full-disk encryption and EDR, since history is stored locally.
- Review on a schedule. Revisit settings as Anthropic ships new controls.
Final Thoughts
Cowork can be deployed safely, but not with the defaults. The pattern is the same as for any privileged system: tight identity, minimal permissions, a controlled network, and logs you can act on. The one thing that has shifted this year is audit, where Anthropic's own coverage has caught up for Enterprise customers. The rest still depends on how you configure it and what you put around it.
Sources: Anthropic: Compliance API coverage for Cowork and Claude Code · Anthropic: monitor Cowork activity with OpenTelemetry · PromptArmor: implement Claude Cowork securely · Cloud Security Alliance: top 6 Cowork security risks
.
TrueFoundry AI Gateway delivers ~3–4 ms latency, handles 350+ RPS on 1 vCPU, scales horizontally with ease, and is production-ready, while LiteLLM suffers from high latency, struggles beyond moderate RPS, lacks built-in scaling, and is best for light or prototype workloads.













.webp)
.webp)
.webp)


.png)
.png)
.png)
.webp)












