Skip to content
Now taking design partners.Roadmap →
Sekeye

Remediation

Findings you can actually close.

A finding about an AI agent is worth little without the rotation plan attached, because the damage travels through the credentials the agent was holding. Rotation checklists and per-host fix plans that a level-1 analyst can execute from a ticket. This is what ships today, Sekeye hands a human the ordered plan, and the human runs it.

See a rotation plan

01

Per-finding fix instructions

Remove the MCP entry from this config file, pull the agent skill, upgrade to at least version x.y.z, drop the model with the unsafe weight format. Concrete, copyable, versioned with the catalog entry, and scoped to the exact host it applies to.

02

Lockfile pull requests (next)

Opening the lockfile bump for you, opt-in per repo, with nothing written without your approval. Not shipped: today Sekeye hands you the exact upgrade path and you apply it.

03

Credential-rotation checklist

The part that matters most when an agent is involved, because an agent's blast radius is mostly the credentials in its environment. Per exposed host: which token types are plausibly present, npm token, GitHub PAT, cloud CLI creds, SSH keys, LLM provider keys, inferred from tool presence and env var names, never by reading secret values → ordered rotation playbook.

Example

Rotation playbook for a host running a compromised MCP server.

Ordered, defensible, and tuned to the tokens Drig inferred were present on that host from tool presence and environment variable names. Never by reading secret values. The same playbook shape covers a package wave like Shai-Hulud; the ordering changes with the blast radius.

  1. 01Rotate the LLM provider and inference-gateway keys the MCP server held in its environment.
  2. 02Revoke and rotate cloud CLI credentials (AWS, GCP, Azure profiles present on disk) that the agent's shell could reach.
  3. 03Cycle GitHub PATs and SSH keys stored on the host, plus any npm tokens tied to affected accounts.
  4. 04Force sign-out and re-auth on IdP-integrated apps.
  5. 05Verify no new refresh tokens were issued during the exposure window.

Roadmap · not shipped

Govern, next: stopping it before the rotation call.

Everything above is cleanup after the fact. The next phase moves the work earlier: a decision before an AI surface installs, and a warning the moment one changes underneath you. Drig is read-only today and none of this is live; it is in build, and we would rather you buy the phase that exists.

Next

A decision before it installs

A new AI surface gets a yes-or-no before it lands on the host, instead of a finding after it already has.

Next

Told the moment it changes

Know when an agent's setup shifts underneath you, at the moment of change rather than at the next scan.

Next

Step in, when you choose to

Move from surfacing the risk to stopping it, opt-in, host by host, and reversible.

The dated version of this sequence lives on the roadmap.