Exposed application credentials: Microsoft sees an agentic framework search GitHub for OAuth and cloud application secrets and then sign in as the applications from 2026-10-28, and Todyl sees an operator use MSP applications to reach 27 organisations, government clients among them
Entra app-credential abuse: agents hunt GitHub for secrets, an operator reaches MSP clients, government among them
Defender actions
- In every Entra tenant you run or manage for others, list the app registrations that hold client secrets, starting with multitenant, delegated-admin (GDAP) and Partner Center apps, security-tool integration apps and Azure DevOps service connections, and search the service-principal and non-interactive sign-in logs for the invalid-secret or expired-secret errors AADSTS7000215 and AADSTS7000222 followed by a success for the same app; on any hit remove all client secrets, certificates and federated credentials on that app, including ones you did not add, disable its service principal if needed and rotate everything stored in any Key Vault it could read, because issued tokens stay valid until they expire.
Analysis
Microsoft reports that since August 2026 threat actors have used the agentic offensive framework ARTEX to submit large volumes of queries to public GitHub for secrets (Microsoft Threat Intelligence, 2026-11-05). As early as 2026-10-28 it saw successful service-principal sign-ins from two IP addresses, one a Tor exit node that may carry unrelated activity, across dozens of applications in several tenants, including a vendor-owned multitenant application whose credential was compromised at the source and used to reach multiple downstream tenants (Microsoft Threat Intelligence, 2026-11-05). Microsoft states that a similar campaign involving the same addresses was also reported by Todyl (Microsoft Threat Intelligence, 2026-11-05).
Todyl, from its managed-detection telemetry, describes one operator who since 2026-11-01 has signed in to MSPs' Microsoft 365 and Azure environments as their own applications with stolen credentials that are still valid: about 45 app registrations at roughly 18 organisations were targeted, more than 33 authenticated, and 27 organisations were reached, among them the MSPs' downstream clients in government (Todyl, 2026-11-03). For nearly every credential the operator requests an Azure Resource Manager token, then a Key Vault token, and within minutes applications that had never been used start authenticating; MSP multitenant and delegated-admin apps and Partner Center tokens reach client tenants, one app sweeping 15 client tenants in about 10-second loops (Todyl, 2026-11-03). Todyl says no Microsoft vulnerability is involved, that it saw no persistence and does not know how the credentials were obtained, and that several seed secrets had only ever been used from staff devices and office networks, which points to a file store or live backend rather than configurations crawled from the internet (Todyl, 2026-11-03). After its detections fired on 2026-11-03 the operator switched to spoofed browser user agents and replayed users' refresh tokens, all blocked by Conditional Access (Todyl, 2026-11-03).
Triage: the invalid-secret and expired-secret errors alone only show that someone holds an old or wrong credential for the app; Microsoft treats a failure followed by a success as the urgent case (Microsoft Threat Intelligence, 2026-11-05).
Cited evidence
Starting in August 2026, Microsoft observed threat actors using an agentic AI offensive framework to submit large volumes of queries to GitHub for secrets.
Sign-in errors AADSTS7000215 (invalid secret) or AADSTS7000222 (expired secret) indicate someone holds a credential for your app; a failure followed by a success is urgent.
Through those apps the actor has reached 27 organizations so far, including the MSPs' downstream clients in government, energy, food and agriculture, transportation, construction and engineering, legal, education and nonprofits.
We have not tied the activity to a specific agentic workflow or offensive framework, and our telemetry does not show an AI model directing it, so we hold that possibility at low confidence.
Sources2
AI-generated · no human review · this permalink is the shareable record for the finding · verify operationally critical claims against the linked primary source.