Skip to content
    Gaincontrol

    AI agents are out of control.

    Gaincontrol finds the AI agents in your code and governs what they're allowed to do: Quin inventories them, and Zino authorizes each call before it runs.

    Simulated fleet

    –NORMAL
    –SCANNING
    –FLAGGED
    –BLOCKED

    The scale of the problem

    Each figure feeds the next: agents run that nobody registered, they hold more access than they need, and the bill lands after the breach.

    1. 82%

      of organizations found an AI agent or workflow that security/IT didn't know existed

      Quin finds them in your repositoriesCloud Security Alliance & Token Security, 2026
    2. 97%

      of non-human identities carry excessive privileges

      Zino scopes each agent's identityEntro Security, 2025 State of NHI Report
    3. $4.63M

      average cost of a breach involving shadow AI, vs. $4.44M baseline

      Zino can revoke a rogue agent in secondsIBM Cost of a Data Breach, 2025

    Agentic AI incidents

    A system prompt is advice. A permission is a control.

    In April 2026 a coding agent working in a startup's staging environment found an account-wide API token in an unrelated file and used it to delete the production database, backups included, in nine seconds. Here is what happened, beat by beat, and the control that applies at each step.

    1. Apr 24

      An agent finds a root key

      A Cursor agent was doing a routine task in PocketOS's staging environment. In a file unrelated to that task it found a Railway API token created only to manage custom domains, but provisioned with account-wide access. As founder Jer Crane put it: "Didn't give it access, it found it."

      Inventory: know what the agent can reach

      Quin
      What happened
      Nobody had a record that a long-lived, account-wide token sat where the agent could read it.
      With the control in place
      The agent, and the credentials, endpoints and tools its code and config can reach, are inventoried, so a root-equivalent token within reach is flagged before anything runs.
    2. Apr 24

      The agent decides to fix something

      It hit a credential mismatch and, on its own initiative, decided to fix it by deleting a Railway volume, assuming the deletion would stay in staging. It didn't check the volume ID or whether the volume was shared across environments.

      Authorization at the moment of the call

      Zino
      What happened
      Nothing stood between the agent's guess and the infrastructure API.
      With the control in place
      Every tool and network call is checked against the agent's policy before it runs. A staging task has no grant to delete infrastructure, so the call is denied and logged.
    3. Apr 24

      One API call, nine seconds

      The agent sent a single volumeDelete request against the production volume. There was no confirmation step, and the volume's backups went with it. "It took 9 seconds," Crane wrote.

      No long-lived credentials in the agent's hands

      Zino
      What happened
      The token was effectively root for the whole account, so the API honoured the request.
      With the control in place
      The agent never holds a long-lived key. Each call gets scoped, per-call access from a credential handler, so a staging task can't borrow account-wide rights.
    4. Apr 25–26

      The agent's confession isn't a record

      Asked what happened, the agent wrote a confession: "I violated every principle I was given." PocketOS restored from a three-month-old backup while a support ticket went unanswered for more than a day. Two days after the deletion, Railway's CEO stepped in and the data was recovered.

      An audit trail the agent doesn't write

      Zino
      What happened
      The clearest account of what ran came from the agent describing itself.
      With the control in place
      Every call, and the authority it ran under, is in a tamper-evident log outside the agent, so you know exactly what was touched while you recover.
    5. Apr 29–May 1

      The fix is scoped access and enforced policy

      Railway made API volume deletions recoverable for 48 hours and called for tokens scoped to the task. Crane's own lesson: "System prompts are advisory, not enforcing." Both point at the same two controls: a scoped identity, and policy enforced outside the prompt.

    1 / 5

    Sources: The Register, Decrypt, Railway

    Frequently asked questions

    All posts
    • · 3 min read

      Authenticating the agent isn't authenticating the action

      "Which agent is this" and "is this action allowed right now" are different questions. Most systems built for agent authentication only ever answer the first one, and treat that as settling the second.

    • · 3 min read

      What a tamper-evident audit trail actually has to prove

      Most "audit logs" built for AI agents are mutable, editable by whoever already has database access, and easy to doctor without leaving a trace. Tamper-evidence is a specific, checkable property, not an adjective a vendor can just claim.

    • · 3 min read

      MCP poisoning: when the server you trusted isn't the one answering today

      An MCP allowlist only ever verified a name at one point in time. If whatever answers to that name changes — a tampered registry entry, a lookalike server, a compromised operator — the allowlist entry still matches, and nothing says so.

    Start with the agents you already have.

    Scan a repository with Quin today, free and open source. When you're ready to govern what it finds, our engineers will walk you through Zino on your own stack.