Three stories landed this week that share a single thread: identity - human, non-human, and agent - is the attack surface that connects every phase of modern intrusions. An AI agent campaign that funds itself by stealing API keys. A device-code phishing technique that compromises AI coding assistants. A Russian APT that turns one cloud workspace breach into a second organization’s compromise. The common denominator is not malware. It is credentials, tokens, and trust.

In the News

AI Agents Steal API Keys, Then Use Them to Hunt for More Keys

Amazon researchers published findings on a campaign that represents a new operational model for credential theft: self-funding AI-agent reconnaissance.

The attack chain starts with internet-facing AI gateway applications - LiteLLM, Dify, and One API are the confirmed targets - running with default credentials or no authentication at all. These gateways are proxies that route requests to AI providers (OpenAI, Anthropic, Cohere) and therefore store provider API keys in their configuration. Attackers’ automated agents scan for these instances, exfiltrate the stored API keys, then immediately use those stolen keys to power the AI models that discover the next batch of exposed gateways.

The recursion is the point. Each stolen key reduces the attacker’s operational cost to zero while extending the campaign’s reach. The models are not being used to generate malware or phishing lures - they are being used to run the reconnaissance loop itself. This is NHI credential sprawl as an active exploit path, not a theoretical risk.

The blast radius extends beyond the gateway operator. Provider API keys often carry billing access, usage logs, and in some cases the ability to fine-tune or query models that hold sensitive training data. A single exposed LiteLLM instance with an OpenAI key becomes a pivot point into the provider’s ecosystem.

What defenders should do: Inventory every AI gateway, proxy, and orchestration layer in your environment. Confirm that no provider API keys are stored with default or no authentication. Implement secrets rotation on a schedule that assumes keys are compromised. Treat AI infrastructure with the same NHI governance rigor applied to service accounts in Active Directory - because attackers already do.

Device-Code Phishing Now Targets AI Coding Agents

Microsoft’s security team published a proof-of-concept demonstrating that device-code authentication - the OAuth flow where a user approves a short alphanumeric code on a separate device - extends directly to AI coding agents like GitHub Copilot CLI.

The attack is straightforward. The attacker initiates a device-code flow against Microsoft’s OAuth endpoint, receives a user code, and delivers that code to the victim developer via a phishing lure (email, chat, or fake error message). The developer navigates to the legitimate Microsoft login page, enters the code, and approves the session. At that point, the attacker holds a persistent OAuth token with the developer’s permissions: code execution, shell access, repository read/write, and in some cases organizational resource access.

No malware is involved. No credential is stolen in the traditional sense. The developer authenticated on a real Microsoft page. The session token - not the password - is what the attacker captures. This is the same device-code phishing technique that targeted Microsoft 365 users in 2024-2025, now extended to AI agent tooling where the OAuth scopes are broader and the session lifetimes are longer.

What defenders should do: Restrict or disable device-code authentication flows via conditional access policies. Enforce phishing-resistant passwordless authentication (FIDO2) for all developer tooling. Monitor for anomalous OAuth token grants - particularly tokens with broad scopes issued through device-code flows. MITRE ATT&CK: T1528 (Steal Application Access Token), T1550.001 (Application Access Token).

COLD RELIC Pivots From One Google Workspace Breach Into a Second US Org

Russia-linked APT COLD RELIC compromised a US organization’s Google Workspace tenant, then used the stolen data - email threads, shared documents, contact relationships, and organizational context - to pivot into a second, unrelated US organization.

The initial access was credential theft against the first victim. The pivot required no exploit, no vulnerability, and no malware delivery. COLD RELIC operated entirely within the trust layer: they replied to real email threads, referenced real shared documents, and leveraged the social capital of a legitimate business relationship. The second organization had no reason to distrust communications from an apparently valid partner.

This is the operational reality of cloud-identity compromise in 2026. The credential is the initial access vector. The workspace data is the reconnaissance. The trust relationship is the lateral movement path - not between machines on a network, but between organizations in a cloud-first supply chain. Traditional network-level detection is irrelevant when the pivot happens through a forwarded email thread.

What defenders should do: Implement ITDR with cross-tenant anomaly detection - watch for unusual email forwarding rules, new OAuth app consent grants, and login patterns from unexpected geographies. Review external sharing policies and conditional access rules for collaboration with partner organizations. Enforce phishing-resistant authentication on your own tenant. MITRE ATT&CK: T1199 (Trusted Relationship), T1114.002 (Remote Email Collection).

Java4Shell and Puppet RCE - Identity Infrastructure in the Blast Radius

Two critical CVEs this week put identity-adjacent infrastructure directly at risk.

CVE-2026-114938 (Java4Shell, CVSS 9.2): A pre-authentication RCE in OpenJDK Java 17.0.6+ on x86-64 systems with AVX-512 instruction sets. Cisco Talos published the technical analysis. CISA added it to the KEV catalog. The confirmed vulnerable products include Keycloak - widely used for SSO, OIDC, and SAML federation - along with Atlassian Jira and Jenkins. For identity teams, the Keycloak exposure is the priority: a pre-auth RCE in your federation layer is a direct path to session forgery and tenant compromise. Patch Java runtimes to 17.0.14+ or 21.0.6+ immediately. Okta has acknowledged the vulnerability and is investigating impact on its own Java-based services.

CVE-2026-115731 (Puppet RCE, CVSS 9.4): A critical RCE in Perforce Puppet with confirmed active exploitation. Puppet manages machine identity at enterprise scale - certificates, configuration, secrets distribution, service-account provisioning. RCE at the Puppet server layer means an attacker can push arbitrary configuration to every managed node, compromising signing keys, certificates, and credentials fleet-wide. Patch immediately and audit all managed-node certificates and keys for signs of tampering.

Defender Action Items

  • Inventory and secure AI gateways: Scan for internet-exposed LiteLLM, Dify, One API, and similar AI proxy instances. Remove default credentials, enforce authentication, rotate all provider API keys stored in these services.
  • Restrict device-code OAuth flows: Disable device-code authentication via conditional access for all developer and AI-agent tooling. Enforce FIDO2 passwordless authentication.
  • Review cross-tenant trust: Audit email forwarding rules, OAuth app consent grants, and external sharing policies in Google Workspace and Microsoft 365. Alert on new external collaborators added to sensitive workspaces.
  • Patch Java4Shell (CVE-2026-114938): Prioritize Keycloak, Jenkins, and Jira instances. Upgrade Java to 17.0.14+ or 21.0.6+.
  • Patch Puppet (CVE-2026-115731): Patch Puppet Server and audit all managed-node certificates and signing keys.

Detection Queries

Monitor for anomalous device-code OAuth token grants - a high-fidelity signal for the Copilot CLI phishing technique Microsoft described. This KQL query surfaces device-code flow completions with unusually broad scopes or from unexpected user agents:

SigninLogs
| where AuthenticationProtocol == "deviceCode"
| where AppDisplayName has_any ("GitHub", "Copilot", "CLI", "Visual Studio")
| where ResultType == 0
| extend ScopeCount = array_length(split(AuthenticationProcessingDetails, ","))
| where ScopeCount > 3 or UserAgent !has "Microsoft"
| project TimeGenerated, UserPrincipalName, AppDisplayName, IPAddress, Location, UserAgent, ResourceDisplayName
| sort by TimeGenerated desc

This catches device-code authentications against developer tooling that succeed with broad scope grants. False positive rate is low in environments where device-code auth is restricted by policy - any hit in a restricted environment is worth investigating.

References


Subscribe to The Identity Brief

Get The Identity Brief in your inbox (Mon/Wed/Fri) - Human, machine, and AI identity security — NHI, ITDR, and the IAM market.