weir
Every credential has one home.
Stop AI agents from leaking your API keys, and prove they didn’t. Free for developers. Proof for organizations.
Ido Goldstein · Founder
OPENAI_API_KEY → api.openai.com, nowhere else
The problem
Your AI agents can see your API keys, and they leak them
Into logs, into prompts sent to the model, into other services. Today nobody can prove they didn’t.
Real example
MLflow, downloaded 20 million times a month, quietly copied developers’ API keys into its monitoring data. Weir caught it with a planted fake key. MLflow fixed it and credited me: CVE-2026-116691.
Source: MLflow security advisory GHSA-j7qq-vr8c-xvw3 · downloads: pypistats, Oct 2026
Why today’s answers fall short
Guessing doesn’t protect keys
AI guarding AI
Adaptive attacks bypassed 12 published AI defenses, most of them more than 90% of the time.
Pattern filters
They hide things that aren’t keys and break output, or miss a key that doesn’t look like one.
Laptop vaults
They protect one developer. Nobody shows a team that every agent, in CI and production, kept its keys home.
Keys don’t need a guess. You know them, so check for them exactly.
Sources: arXiv 2510.09023 · github/copilot-cli #4065
Weir · open source · free for every developer
Every credential has one home
A key may go only to its own service. Anywhere else is a leak, caught by exact match, with no AI in the loop.
Don’t fetch
Block the commands that hand out keys
Don’t show
Strip your exact keys from what the model reads
Don’t send
The agent holds a decoy; a gate holds the real key
Prove it
Plant fake keys, run the agent, check every exit
On PyPI today: weir-scan and weir-run · github.com/IdoGol24/weir
For organizations · paid
Proof across every agent, from CI to production
Test in CI
Signed result
Deploy gate
Live alarm
One view
Every agent and deployment: verified, protected, alarms.
Audit-ready evidence
Signed results and reports mapped to AIUC-1 and SOC 2.
Alarms and response
A leaked decoy blocks the call, alerts the team, and can roll back.
Free: the tools. Paid: the view, the evidence and the alarms, once agents reach production or an audit begins.
Evidence only, never secrets · AWS first: ECS, EKS and CI
Ido Goldstein, headshot
Ido Goldstein, founder
7 years building distributed systems and security at Meta, Cyber Startups and the IDF. Credited reporter of CVE-2026-116691. Creator of Weir.
The ask
Looking for design partners
Teams running AI agents on AWS: I run the same test that found the MLflow leak, in your CI. A free two-week pilot, a signed report, and nothing leaves your environment.
In return
  • Your feedback every two weeks
  • A logo or a quote
  • Permission to share anonymised results
idogold@ralis.ai · github.com/IdoGol24/weir