· via dev.to (home feed)
crewAI, OpenClaw and LiteLLM advisories show agent guardrails failing open
Three security advisories published on 19 and 20 November by crewAI, OpenClaw and LiteLLM show agent guardrails — read-only SQL, tool policy, key isolation — failing open on edge cases their authors did not anticipate.

Three advisories in two days
On 19 and 20 November, three pieces of the AI agent toolchain published security advisories within about 24 hours of each other: crewAI, OpenClaw and LiteLLM. According to an analysis on dev.to by Humanbound, a vendor of agent security testing, the common thread is not a shared bug but a shared pattern: in each case a guardrail existed, was switched on, and failed open on an edge case its authors had not anticipated.
The advisories are GHSA-wp6h-6xxh-g8fp for crewAI, GHSA-pw3x-g8v8-fvj2 for OpenClaw, and GHSA-4xm2-x5g5-wc8q for LiteLLM.
crewAI: three hardened components, three bypasses
The crewAI advisory, rated high and published 19 November, stemmed from a review of version 1.15.30 and covers bypasses in components that had already been hardened.
The first involves deserialization. crewAI ships the CREWAI_DESERIALIZE_CALLBACKS environment variable so untrusted checkpoints and serialized payloads cannot import arbitrary Python modules. But BaseTool deserialization had a second path that skipped the guard: a dotted path in cache_function was resolved into a live callable that runs during tool execution, and an unknown tool_type was passed straight to importlib.import_module() with no trust check and no verification that the result was actually a BaseTool.
The second involves NL2SQLTool, which runs read-only by default (allow_dml=False) and tries to block writes hidden inside CTEs or EXPLAIN ANALYZE. Per the write-up, three constructs slipped past the validator: WITH cte AS MATERIALIZED and AS NOT MATERIALIZED did not match the parser's regex, so validation returned without raising, failing open; a leading comment inside a CTE body, as in WITH d AS (/*comment*/ DELETE FROM users RETURNING *) SELECT * FROM d, hid the DELETE from the check; and EXPLAIN (ANALYZE true) DELETE FROM users was not recognised as an ANALYZE form, so the delete ran. Denylisting a language as rich as SQL is fragile at the best of times, and for agents it is worse because the SQL is written by a model that can be steered by whatever text it last read.
The third involves file path validation. Symlinks resolving outside the working directory were rejected, but the comparison used a string prefix. With a working directory of /app/work, a link to /app/work_secrets/key.pem passed. Only the final path component was checked, so a symlinked parent directory was never examined at all.
The first two issues are fixed in crewAI 1.15.31; the path issue is fixed in crewai-files 1.15.44, which now uses Path.is_relative_to on the resolved path each time. The advisory reportedly offers no workaround beyond avoiding untrusted paths and symlinks.
OpenClaw: delegation skipped approval
Also published 19 November and rated high, the OpenClaw advisory is brief and pointed. In affected versions, the openclaw tool could apply persistent configuration changes without approval, even when tool policy denied both execution and file writes for the agent in question. The consequence, per the advisory as cited by Humanbound: a restricted agent could enable host command execution through configuration changes. It could not run commands directly, but it could edit the settings that decide whether it may.
Fixes landed in stable 2026.11.4 and extended-stable 2026.10.33. Until an upgrade is in place, the advisory recommends setting tools.exec.mode to ask, denying untrusted delegated changes, keeping allowlists narrow, and not sharing a single Gateway between users who do not trust each other.
LiteLLM: key isolation at the gateway
Published 20 November and rated medium, the LiteLLM advisory affects deployments that use virtual keys with no assigned user — the service-account pattern common in agent setups. Such a key could call /key/info and read the stored records of any other userless key, including keys belonging to other teams: alias, team, allowed models, budgets, spend and metadata.
The advisory is explicit about the limits. Raw keys and hashes were not exposed, and the bug does not allow using, changing or deleting another key. Admins, user-assigned keys and same-team reads are unaffected. Even so, budget and model metadata is useful reconnaissance for anyone planning to abuse a gateway.
What to do now
Upgrade: crewAI to 1.15.31 with crewai-files 1.15.44, OpenClaw to 2026.11.4 or 2026.10.33, and LiteLLM per its advisory. Beyond patching, the Humanbound write-up suggests tests that generalise: probe every "read-only" tool with valid but unusual syntax such as comments and alternate keywords; check whether restricted agents can delegate around their restrictions via config tools or other agents; feed tools symlinks, symlinked parent directories and prefix-colliding names; and check what one agent's credentials can read about other teams' keys and budgets.
Why it matters
Each of these guardrails promised a hard boundary — no writes, no arbitrary imports, no command execution, tenant isolation — and each was enforced in one place while assumed everywhere else. That failure mode is more dangerous for agents than for ordinary software because agents explore: they chain tools, retry, rewrite their own SQL and delegate to other agents. As Humanbound frames it, an attacker using indirect prompt injection does not need to find the edge case themselves; they can let the agent search for it. If your agent's safety depends on a flag that says "read-only" or "restricted", that flag needs adversarial testing just as much as the model does.
- #ai-agents
- #security
- #crewai
- #openclaw
- #litellm