Read-only, restricted, contained: the agent guardrails that failed open.
Humanbound's engine and CLI are open source: github.com/humanbound/humanbound On 19 and 20 November, crewAI, OpenClaw and LiteLLM each published a security advisory. Every one involves a guardrail that existed and was s
Humanbound's engine and CLI are open source: github.com/humanbound/humanbound
On 19 and 20 November, crewAI, OpenClaw and LiteLLM each published a security advisory. Every one involves a guardrail that existed and was switched on: a read-only SQL mode, a deny policy for restricted agents, symlink containment, per-team key isolation. Each one failed on an edge case its authors didn't anticipate. If your agent's safety depends on a flag saying "read-only" or "restricted," that flag needs adversarial testing just like the model does.
crewAI: three hardened components, three bypasses
Published 19 November and rated high, GHSA-wp6h-6xxh-g8fp comes from a review of crewAI 1.15.30. The advisory itself describes all three issues as bypasses in "previously hardened components."
1. A deserialization guard with a side door
crewAI added the CREWAI_DESERIALIZE_CALLBACKS environment variable so untrusted checkpoints or serialized payloads can't import arbitrary Python modules. But BaseTool deserialization had its own path that skipped the guard:
- A dotted path in
cache_functionwas resolved into a live callable even with the guard disabled, and that callable runs during tool execution. - An unknown
tool_typewas passed straight toimportlib.import_module()with no trust check and no check that the result was actually aBaseTool.
The guard was real. It just wasn't the only way in.
2. "Read-only" SQL that could still delete
NL2SQLTool runs in read-only mode by default (allow_dml=False) and tries to block writes hidden in CTEs or EXPLAIN ANALYZE. The advisory lists three ways valid SQL slipped past the validator:
-
WITH cte AS MATERIALIZED (...)andAS NOT MATERIALIZED (...)didn't match the parser's regex, so validation returned without raising: it failed open. - A leading comment inside a CTE body, as in
WITH d AS (/*comment*/ DELETE FROM users RETURNING *) SELECT * FROM d, hid theDELETEkeyword from the check. -
EXPLAIN (ANALYZE true) DELETE FROM userswasn't recognized asANALYZE, so the delete ran.
This is the classic problem with denylisting a language as rich as SQL. For an agent, it's worse: the SQL is written by a model that can be steered by whatever text it last read.
3. Symlink containment checked by string prefix
FilePath validation rejected symlinks that resolved outside the working directory, but it compared paths with startswith. With a working directory of /app/work, a link to /app/work_secrets/key.pem passed. It also only checked the final path component, so a symlinked parent directory was never checked at all.
Fixes: issues 1 and 2 are fixed in crewAI 1.15.31; issue 3 in crewai-files 1.15.44, which now uses Path.is_relative_to on the resolved path every time. The advisory notes there is no workaround other than avoiding untrusted paths and symlinks.
OpenClaw: delegation skipped approval
Also published 19 November and rated high, GHSA-pw3x-g8v8-fvj2 is short and pointed. In affected versions, the openclaw tool could apply persistent configuration changes without approval, even when tool policy denied execution and file writes for that agent.
The impact, in the advisory's words: this "could let a restricted agent enable host command execution through configuration changes." In other words, the agent couldn't run commands directly, but it could change the settings that decide whether it runs commands.
Fixes: stable 2026.11.4, extended-stable 2026.10.33. Until you upgrade, the advisory recommends setting tools.exec.mode to ask, denying untrusted delegated changes, keeping allowlists narrow and not sharing one Gateway between users who don't trust each other.
LiteLLM: tenant isolation at the gateway
Published 20 November and rated medium, GHSA-4xm2-x5g5-wc8q affects deployments that use virtual keys with no assigned user, such as service-account keys. Such a key could call /key/info and read the stored record of any other userless key, including keys belonging to other teams: alias, team, allowed models, budgets, spend and metadata.
The advisory is clear about the limits. Raw keys and hashes were not exposed, and the bug doesn't allow using, changing or deleting another key. Admins, user-assigned keys and same-team reads aren't affected. Still, many agent deployments hand each agent a userless service key, and budget and model metadata is useful reconnaissance for anyone planning to abuse a gateway.
The pattern: guardrails as labels
Put the three side by side:
| Guardrail | What it promised | How it failed |
|---|---|---|
crewAI allow_dml=False
|
No writes | Valid SQL the parser didn't recognize |
| crewAI symlink check | Stay in the working dir | String prefix match, parent dirs unchecked |
| crewAI deserialize guard | No arbitrary imports | A second code path skipped it |
| OpenClaw tool policy | Restricted agent can't exec | Agent changed the config that grants exec |
| LiteLLM virtual keys | Teams isolated | Userless keys shared one visibility scope |
None of these is exotic. Each is a boundary that was enforced in one place and assumed everywhere else. That matters more for agents than for normal apps because an agent explores. It chains tools, retries, rewrites its own SQL and delegates to other agents. An attacker steering it through indirect prompt injection doesn't need to find the edge case themselves. They can let the agent search for it.
What to test in your own stack
- Write probes against every "read-only" tool. Use valid syntax the guard might not expect: comments, alternate keywords, nested statements.
- Delegate around your restrictions. Can a restricted agent ask another agent, or a config tool, to do what it can't?
- Feed tools paths you don't control. Use symlinks, symlinked parent directories and names that share a prefix with allowed directories.
- Check what one agent's credentials can see about the others. Look at other teams' keys, budgets and model lists, not just whether it can make requests.
-
Upgrade: crewAI to 1.15.31 (and
crewai-files1.15.44), OpenClaw to 2026.11.4 or 2026.10.33, and LiteLLM per its advisory.
Find out where your guardrails fail open. Humanbound's free Community plan (€0) runs hosted attack models against your agent, probing tool misuse, prompt injection and boundary bypasses like these. No agent ready yet? Try the built-in FinAssist demo first. Sign up at app.humanbound.ai.
References
- crewAI: GHSA-wp6h-6xxh-g8fp (published 19 November 2026)
- OpenClaw: GHSA-pw3x-g8v8-fvj2 (published 19 November 2026)
- LiteLLM: GHSA-4xm2-x5g5-wc8q (published 20 November 2026)
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.