International
The Morning Dev

Only the news worth knowing

‹
 
B42/100

Security incident · checklist estimated

Read-only, restricted, contained: the agent guardrails that failed open

Tier B · Patch this week – Put it in the sprint

Tier B because: Full takeover OR secrets exposed OR supply chain

Reach

Attack

Damage

Fix

Full takeover

The Brief

  • On 19 and 20 November, crewAI, OpenClaw and LiteLLM each published a security advisory.
  • Each advisory involves a guardrail that existed and was switched on, such as a read-only SQL mode or a deny policy.
  • Humanbound's engine and CLI are open source, at github.com/humanbound/humanbound.

Why it matters

Enabling a built-in guardrail did not reliably contain agent behaviour in these frameworks, so developers should not rely on a single switch to restrict what an agent can do.

The checklist behind the tier

Checklist · 2 of 10 yes · estimated from older data, not yet checked against the sources
YesHits tech teamsDoes it affect software, devices or services that developers, ops or IT teams build with, run or are responsible for?
NoCore infrastructureIs the flaw in, or the compromise of, software that nearly every team runs?
NoPopular toolIs the flaw in, or the compromise of, a widely used framework, tool, platform, OS or package?
NoExploited in the wildDo sources report real attacks using it?
NoPublic exploitIs a working exploit or detailed proof-of-concept public?
NoRemote, no loginCan an attacker use it over the network without credentials or user interaction?
YesFull takeoverDoes it lead to code execution, root/admin rights or full account takeover?
NoSecrets exposedWere credentials, tokens, keys, configs, source code or customer data stolen or exposed?
NoSupply chainWas malicious code shipped through a channel developers trust?
NoNo fix yetIs there no patched version yet?

Timeline

  1. crewAI, OpenClaw and LiteLLM published security advisories (19 and 20 November)

  2. DEV Community: ai

    Read-only, restricted, contained: the agent guardrails that failed open.

Source

Read-only, restricted, contained: the agent guardrails that failed open. ↗ DEV Community: ai · 44 min ago

First seen · 1 source · Full story page · How tiers are decided