CVE-2026-102610evalstate fast-agent OS command injection
CVE-2026-102610 is a OS command injection in evalstate fast-agent, rated medium (6.4 on CVSS 3.1).
Why this priority: Medium severity, with no exploitation signal.
Published · Updated · NVD status: Received
What it is
Input reaches an operating-system command line, letting an attacker run their own commands on the host. A function that should require sign-in can be called by anyone.
How it is reached
An attacker can exploit it over the network, with an administrator account, but someone has to act first, such as opening a file or following a link. Success also depends on conditions the attacker does not fully control.
What an attacker gains
A successful attack can read the data it protects, change data or behaviour and take it offline. In practice, this kind of flaw lets an attacker run commands or code with the privileges of the affected service.
Is it being exploited?
There is no public report of exploitation yet.
What to do
- Upgrade evalstate fast-agent to 0.7.21 or later.
- Until it is patched, review who holds accounts that can reach it.
- Warn the people who use it, and filter files and links on the way in.
Description from NVD
fast-agent is a framework for building, running, and evaluating AI agents with Model Context Protocol support. Prior to 0.7.21, fast-agent serve uses HTTP transport bound to 0.0.0.0 on port 8000 without authentication, allowing any network client that can reach the /mcp endpoint to invoke the exposed agent tool. When the server is started with --shell, the agent can call the local execute tool implemented by tools/local_shell_executor.py, so an unauthenticated remote client can cause operating-system commands to run as the server user; without --shell, the client can still consume the operator's model and configured MCP tools. The defaults are defined in cli/commands/serve.py and mcp/server/agent_server.py, and the non-interactive server permission prompt does not prevent the described access. This issue is fixed in version 0.7.21.
What this means for your architecture.
A CVE names a product. Whether it matters depends on where that product sits in your systems: what reaches it, which trust boundary it guards and what it can reach once compromised. Veriom maps your components, their relationships and your trust boundaries from your code, cloud and delivery evidence, so you can see where a component like this sits. The radar itself does not look at your systems.
Components it touches
- Application servers and the hosts they run on. The flaw sits between untrusted input and code or command execution.
- Template engines. The flaw sits between untrusted input and code or command execution.
- Sign-in, session and token handling. The flaw sits between anonymous callers and authenticated sessions.
- Identity providers and SSO. The flaw sits between anonymous callers and authenticated sessions.
Trust boundaries in play
- Authenticated network access. An account is needed, so exposure follows who can sign in: customers, partners or staff.
- People who open files and links. Mail, chat and browsers are the delivery path.
- Between untrusted input and code or command execution. An attacker who crosses it can run commands or code with the privileges of the affected service.
- Between anonymous callers and authenticated sessions. An attacker who crosses it can act as a signed-in user, or an administrator, without valid credentials.
Questions to answer about your own systems
- Where does evalstate fast-agent run in your systems, and at which version?
- Which of those instances can be reached from the internet or partner networks?
- What can it reach once compromised: which data stores, secrets and services?
Sources
- CVE records and CVSS scores from the NVD CVE API. This product uses the NVD API but is not endorsed or certified by the NVD.
- Exploitation status from CISA’s Known Exploited Vulnerabilities catalog.
- Exploit likelihood from FIRST’s Exploit Prediction Scoring System (EPSS), refreshed daily.
- Explanations are written by Veriom from fixed templates over each record’s data, not by a person reviewing each CVE. Descriptions are quoted from the records as published.