Agent tool calls are getting a gatekeeper, after the exploits
MCP command injection runs with the developer's own access, and a LiteLLM sandbox escape ran as root. This week the fix arrived as gateways and egress checks.
- CVE-2025-6514 affected mcp-remote versions 0.0.5 to 0.1.15 and was rated with a CVSS score of 9.6. m:…2369
- LiteLLM's guardrail test endpoint allowed user-supplied Python code to escape its sandbox and run as root in the default Docker image; CISA added CVE-2026-40217 to the KEV catalog on 2026-11-18, and the fix is LiteLLM version 1.83.11 or later. m:…6557
- LangChain launched LangSmith Tool Gateway in public beta, with deny policies, human-in-the-loop approval for sensitive tool calls, and audit logs of which tool ran and as whom. Self-reported m:…9103
- In a preprint, the PIPETTE red-teaming method produces defenders that achieve near-zero attack success rates on static attacks while keeping benign task utility close to the undefended baseline. Self-reported m:…5061
The attack runs as you
The attack that matters for agent tool use is not a clever prompt. It is a shell command. MCP command injection happens when an MCP server or client passes untrusted input into an operating system command, and Airia notes it is distinct from prompt injection, though the two are often chained together in practice. The exposure is wide because local MCP servers often run with the developer's full system access. m:…2369
The pattern already has a named case. CVE-2025-6514 affected mcp-remote versions 0.0.5 to 0.1.15 and was rated with a CVSS score of 9.6. Airia says developers should avoid using shell commands to open URLs in MCP clients, and that running servers in containers or isolated environments mitigates the risk. Its own gateway lets admins assign access at the app, tool, and parameter level, and its detection blocks prompt injection with sub-300ms latency. m:…2369
One layer up, the same failure
The LiteLLM case shows the same failure one layer up. LiteLLM's guardrail test endpoint allowed user-supplied Python code to run in a sandbox that could be escaped by a proxy administrator, and the escaped code ran with the privileges of the proxy process, which is root in the default Docker image. CISA added CVE-2026-40217 to the KEV catalog on 2026-11-18. The flaw is rated CVSS v3.1 8.8 High and CVSS v4.0 7.5 High, and the fix replaces the regex-based sandbox with RestrictedPython in LiteLLM version 1.83.11 or later. The federal remediation deadline is 2026-12-02. m:…6557
The fix is a gate
The vendor response this week is a gate, not a warning. LangChain launched LangSmith Tool Gateway in public beta to manage credentials, identity, and access for agent tool calls, so teams can grant agents access to tools without hand-rolling auth. Admins can set deny policies at the org or workspace level, approval policies allow human-in-the-loop confirmation for sensitive tool calls, and audit logs capture which tool ran and as whom. The gateway works with MCP clients including LangChain, Agents, Claude Code, and Cursor, and includes built-in servers for Google, Salesforce, GitHub, and Linear. m:…9103
OpenBox takes the same idea to the sandbox exit. Its governance layer for NVIDIA OpenShell sandboxes is a fail-closed policy decision point on the egress path of every sandbox, evaluating model calls, tool calls, and web requests before they leave. If no verdict arrives within the 29-second budget, the request is refused, and risky actions can be held for human approval. The records are tamper-evident through hash-chaining, and the layer cannot be switched off from inside the sandbox. It is available for NVIDIA OpenShell v0.1.8. m:…2961
Research published as a preprint adds a testing discipline behind the gates. In a preprint, PIPETTE uses dictionary learning-based rewards to encourage broader coverage of attack strategies, because earlier red-teaming tools suffered from mode collapse and high attack success rates during red-teaming did not guarantee robustness gains from adversarial training. Its defenders achieve near-zero attack success rates on static attacks while keeping benign task utility close to the undefended baseline. The lesson is that a gate is only as good as the attacks it was tested against. m:…5061
Put together, the week points in one direction: tool use is getting a permission layer whether teams planned one or not. The exploits run as the developer, the sandbox, or root, and the fixes all sit at a choke point: a gateway, an egress check, or a red-team suite. Teams that still let agents call tools directly are not saving time. They are skipping the layer where the risk is now concentrated. m:…2369 m:…6557
What is still unresolved
- Whether the LangSmith Tool Gateway's public beta covers the MCP clients teams already run in production.
- How many deployed LiteLLM proxies still run a version before 1.83.11.
- Whether PIPETTE's near-zero attack success on static attacks holds against live, adaptive attacks.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.