Agent Field Kit @AgentFieldKit

Per LiteLLM, the attacker may have gotten GitHub CLI tokens from logs on an internet-facing internal LLM gateway. Those tokens later unlocked the release workflow. Free check today: search your gateway logs for tokens, purge them, and redact request/response bodies.

Your agent's cost isn't price per token. It's tokens × steps × retries × the runs that fail and get re-run. Measure cost per successful task, or the invoice will measure it for you.

LiteLLM says its source tag for 1.107.0 was clean. The artifacts were built from a lookalike repo, Berri-AI/litellm instead of BerriAI/litellm, using stolen developer credentials. Signed and published through the normal pipeline. A clean repo isn't enough on its own.

agentfieldkit.com/go/x2-afk-pinn…

Agent Field Kit: editable templates for running agents in production. Eval plans, incident runbooks, a vendor security questionnaire, a prompt-injection test pack. From $29:

AFK: away from keyboard. Which is exactly when your agents do the weird stuff.

Oct 21 brought two separate disclosures aimed at the keys AI apps run on: a hosting platform's stored secrets and a poisoned LiteLLM release. We wrote a plain explainer: how each path worked, who's affected, and free steps first.

brb, agents doing a no-hit run Speedrunners practice every split. Your agent should pass the eval set before it ships, and again every time the prompt or model changes. Eval test plan: #AIAgents #LLM

#AIAgents #AISecurity

Prompt-injection test tip: don't only test the chat box. Test every place text enters the agent: retrieved docs, tool outputs, emails, file names, web pages. The agent can't tell your data from an instruction unless you make it.

agentfieldkit.com/go/afk-x-cyber…

Coding agent pulling skills from npm? GitHub now lists versions of agentic-awesome-skills and opencode-skills-collection as malware (GHSA-rgwm-3jj9-h563, GHSA-9855-prwc-mfp2). Run npm ls. If you have a listed version, rotate keys from a clean machine.

agentfieldkit.com/go/afk-x-cyber…

Railway update (Oct 22, 1:29 PM PT): API tokens AND Project Tokens created before the incident stay invalidated, so make new ones for CLI/CI and delete the old CI secrets. Hobby builds aren't back yet. Rotate leaked keys at the issuer anyway. Steps:

agentfieldkit.com/go/afk-x-cyber…

Agent front end on Next.js? 16.4.1 and 15.5.28 fix two critical unauthenticated RCEs (use cache typed arrays; SVG image optimization). GitHub added them to its advisory DB today, so scanners may only now flag them. Upgrade, rebuild, check sharp 0.35.5+.

agentfieldkit.com/go/afk-x-cyber…

Patch list for agent builders this week: - MCP TypeScript SDK: 1.32.0 (or server/client 2.3.0) - vLLM: 0.31.0, and keep trust_remote_code off unless needed - Node.js: 22.23.4 / 24.22.1 / 26.12.1 Upgrade, rebuild images, redeploy. Free first steps:

agentfieldkit.com/go/afk-x-cyber…

Railway update for agents hosted there: Railway has invalidated all Railway API tokens and is rotating storage bucket credentials. Create a new API token for any CI job or script, then redeploy so services pick up the new bucket creds. Free steps:

agentfieldkit.com/go/afk-x-cyber…

LiteLLM 1.107.0 follow-up: Wiz reports stolen AWS role credentials were tested against Bedrock within hours. If a host ran 1.107.0, check CloudTrail for GetCallerIdentity and bedrock:ListFoundationModels from unknown IPs, and add a billing alarm.

Things that make the next rotation shorter: - One credential per agent - References instead of pasted values - An expiry on every key that supports one - A named owner for each secret