Taiwan Information Security Association

AI RiskThu, 22 Oct 2026

⑤ Hermes and OpenClaw: How API and identity flaws exposed government systems

The investigation traces a path from public APIs through accounts and single sign-on to internal services and data.

01

Article

On August 12, Dream Research Labs described a July operation using the Hermes and OpenClaw agent frameworks. The published report did not name the underlying models. Taiwan’s Administration for Cyber Security confirmed on August 13 that the activity combined human operators with AI agents. It urged organizations to check secondary systems such as backups and testing environments.

Dream also observed broader reconnaissance against energy, nuclear security and supplier targets. Some controls interrupted follow-on activity. For example, Forms Authentication blocked access to an uploaded web shell.

Attack flowHow the attack progressed
  1. 01

    Map public entry points

    Attackers analyzed frontend applications, logins and APIs.

  2. 02

    Obtain a usable identity

    Weak passwords, unauthorized APIs and gaps in JWT verification provided paths to access.

  3. 03

    Reach internal services

    Single sign-on extended access across applications, opening opportunities to query or extract data.

What to do: from the perimeter to internal systems

Response flowResponse steps
  1. 01

    Inventory all entry points

    Include production, testing, backup, container and outsourced environments.

    See what to doResources: 1

    TWISA member solutions

    Inventory all entry points

    Include production, testing, backup, container and outsourced environments.

  2. 02

    Validate APIs and identity

    Test with unauthenticated and low-privilege accounts. Check JWT signatures, issuer, audience and expiry.

    See what to doResources: 1

    TWISA member solutions

    Validate APIs and identity

    Test with unauthenticated and low-privilege accounts. Check JWT signatures, issuer, audience and expiry.

  3. 03

    Limit access after SSO

    Enforce role and data permissions at each service, and revoke abnormal accounts and sessions.

    See what to doResources: 1

    TWISA member solutions

    Limit access after SSO

    Enforce role and data permissions at each service, and revoke abnormal accounts and sessions.

  4. 04

    Correlate identity and data events

    Trace API and SSO logs through to data access before restoring affected services.

    See what to doResources: 1

    TWISA member solutions

    Correlate identity and data events

    Trace API and SSO logs through to data access before restoring affected services.

Defense architectureLayered defense

Public and secondary systems, APIs, identity providers and sensitive data

Allow authorized traffic and apply controls at each layer.

  1. 01

    Inventory all entry points

    Public and secondary systems, APIs, identity providers and sensitive data

    Validate the controlInclude production, testing, backup, container and outsourced environments.

  2. 02

    Validate APIs and identity

    Public and secondary systems, APIs, identity providers and sensitive data

    Validate the controlTest with unauthenticated and low-privilege accounts. Check JWT signatures, issuer, audience and expiry.

  3. 03

    Limit access after SSO

    Public and secondary systems, APIs, identity providers and sensitive data

    Validate the controlEnforce role and data permissions at each service, and revoke abnormal accounts and sessions.

Feed logs and traffic data from each layer into monitoring and response.

  1. Correlate identity and data eventsTrace API and SSO logs through to data access before restoring affected services.

TWISA member solutions

Continue the discussion at CIX

A future CIX meetup can use the API-to-SSO path to test a team’s threat model. Bring a secondary application or staging service and walk through the permissions that still apply after sign-in.

Sources and further reading

Column updated: 2026/10/25.