AI agents are no longer theoretical threats to identity infrastructure - they are operational ones. This week, Amazon published captured logs showing AI agents running recursive credential-theft loops against exposed AI gateways, the FBI warned that ARTEX is driving real supply-chain intrusions from MSP tools into customer M365 tenants in under three hours, and GitGuardian quantified what most identity teams already suspected: leaked tokens stay valid because nobody revokes them. The common thread is non-human identity - API keys, JWTs, OAuth tokens, service accounts - and the fact that most organizations have no governance over any of it.

In the News

Amazon Catches AI Agents Running Recursive Key-Theft Loops Against LiteLLM Gateways

Amazon’s security team published live logs from an active campaign in which AI agents systematically scan the internet for exposed LiteLLM, Dify, and One API gateway instances. These gateways are internal AI proxy tools that organizations deploy to manage access to providers like OpenAI, Anthropic, and Cohere. They store provider API keys - and when exposed with default configurations, those keys are trivially extractable.

What makes this campaign different from standard credential theft is the recursion. The attacking agent steals a provider API key from one gateway, uses that key to invoke the AI model it now controls, and directs that model to find the next exposed gateway. The stolen credential powers the tool that steals the next credential. Amazon describes this as the first publicly documented AI-on-AI supply-chain loop.

The identity implications are significant. These provider API keys are non-human identities with direct access to paid AI services, often with no rate limiting, no IP restriction, and no rotation policy. Most organizations deploying LiteLLM internally treat it as a development convenience, not as a production service requiring secrets management and network segmentation. The attackers are treating it as a production target.

What defenders should do: Audit any internal AI gateway deployment for internet exposure. Store provider API keys in a secrets manager (HashiCorp Vault, AWS Secrets Manager) with automated rotation. Restrict gateway network access to authorized internal services only. Monitor provider API usage for anomalous volume or source-IP changes that indicate key theft. Source: The Hacker News

FBI Warns ARTEX AI Agents Automate ConnectWise-to-M365 Supply-Chain Attacks

The FBI issued an advisory warning that ARTEX - the AI-driven attack framework that first surfaced November 13 - is now powering real supply-chain intrusions through ConnectWise ScreenConnect cloud instances. The attack chain is straightforward: ARTEX compromises an MSP’s ScreenConnect deployment, hijacks legitimate remote-support sessions to reach customer endpoints, and deploys a worm Microsoft calls SCKit. SCKit installs unauthorized remote-access clients and harvests credentials from the customer’s M365 environment.

The FBI reports that ARTEX cuts time from initial access to privileged account compromise to under three hours. Microsoft’s Security Blog confirms the SCKit worm deploys via ScreenConnect’s legitimate session-hijacking capabilities - meaning network-level detection of the initial pivot is difficult because the traffic looks like normal MSP remote support.

This is an identity problem at every stage. The initial compromise exploits MSP trust relationships - the ScreenConnect session is a delegated identity with broad access. The lateral movement relies on credential harvesting. The persistence mechanism creates unauthorized OAuth app registrations in M365. Organizations relying on MSPs need conditional access policies that restrict third-party remote-access tools, continuous monitoring for new OAuth app registrations, and ITDR coverage that detects behavioral anomalies in MSP-originated sessions.

What defenders should do: Enforce conditional access policies restricting which remote-access tools can reach M365 resources. Monitor for unauthorized OAuth app registrations in Entra ID. Deploy phishing-resistant authentication (FIDO2) to block the credential-harvesting stage. Audit MSP access scope and session logging. Source: BleepingComputer | Source: Microsoft Security Blog

334,000 Leaked JWTs Still Valid - Token Revocation Is the Missing Control

GitGuardian built 345 attribution rules that matched 86% of 2.4 million leaked JSON Web Tokens back to their issuing services - Google, Auth0, Okta, Azure AD, and dozens of others. When researchers replayed the tokens against live APIs, 75% validated successfully. At the time of detection, 334,000 tokens were still active and usable by anyone who possessed them.

The core issue is that JWT-based authentication systems conflate cryptographic validity with authorization validity. A token signed by a trusted issuer is accepted by any relying party that trusts that issuer’s public key - regardless of whether the token has been compromised. Most SSO deployments issue tokens with multi-hour or multi-day lifetimes and have no revocation mechanism beyond waiting for natural expiry. Some identity providers offer token revocation endpoints, but very few organizations have automated pipelines that detect a leaked token and trigger revocation within minutes.

This is a non-human identity governance gap. JWTs function as bearer credentials - they are, operationally, service account passwords with an expiration date. When one leaks in a Git commit, a CI log, or an error message, the blast radius is a fully authenticated session for the remaining lifetime of the token. GitGuardian’s dataset quantifies what practitioners have long known: the gap between “leaked” and “revoked” is measured in hours to days, not minutes.

What defenders should do: Implement secrets detection in code repositories and CI/CD pipelines that triggers automated token revocation workflows. Enforce short-lived token policies (minutes, not hours). Use token-binding mechanisms where available to tie tokens to specific devices or sessions. Audit identity provider configurations for token lifetime settings. Source: GitGuardian

Talos Documents 700+ Public AI Pentest Consoles - Commodity Attack Automation Arrives

Cisco Talos published research documenting the broader AI-powered attack tool ecosystem beyond ARTEX. The findings include over 700 live PentAGI instances - web-UI AI pentest frameworks - exposed on the public internet, alongside Hephaestus (a Telegram-controlled, single-script deployment) and HexStrike. All three frameworks use Censys and Shodan for automated target discovery, CAPTCHA bypass services, residential proxies for attribution evasion, and hash-cracking-as-a-service for offline credential attacks.

The full stack costs under $100 per month. This is the commoditization inflection point for AI-driven attack automation. The barrier to running a fully automated reconnaissance-to-exploitation pipeline is no longer technical expertise - it is a Telegram account and a payment method.

For identity teams, the implication is velocity. When AI agents automate the discovery of exposed services and the exploitation of default credentials, the window between “exposed” and “compromised” collapses from days to hours. Any internet-facing service with default credentials, weak authentication, or missing MFA is now being found and tested at machine speed.

What defenders should do: Run external attack surface discovery against your own infrastructure on a continuous basis. Eliminate default credentials on all internet-facing services. Enforce phishing-resistant authentication on every externally accessible application. Treat AI pentest tool scanning as the new baseline threat model for internet exposure. Source: Cisco Talos

Defender Action Items

  • Audit AI gateway exposure: Identify any LiteLLM, Dify, or One API instances reachable from the internet. Move provider API keys to a secrets manager with automated rotation. Restrict network access to internal consumers only.
  • Harden MSP trust chains: Enforce conditional access policies in Entra ID / your IdP that restrict third-party remote-access tools. Monitor for new OAuth app registrations originating from MSP sessions. Deploy FIDO2 phishing-resistant auth for all privileged accounts.
  • Implement token revocation pipelines: Configure secrets detection in Git repositories and CI/CD that triggers automated JWT and API key revocation. Reduce token lifetimes to the shortest operationally viable window. Audit identity provider token lifetime settings.
  • Patch CVE-2026-114713 (UnrealIRCd, CVSS 8.6): CISA KEV-listed as of November 15. IRC is seeing renewed use as C2 infrastructure for AI agents.
  • Patch CVE-2026-6460 (PostgreSQL, CVSS 8.8): 22-year-old type-confusion bug allowing authenticated RCE. Prioritize for multi-tenant environments, AI training pipelines, and SaaS backends.

Detection Queries

Detect anomalous OAuth app registrations in M365 environments - a key indicator of the ARTEX/SCKit supply-chain attack:

1-- Splunk SPL: New OAuth app consent in M365 from MSP IP ranges
2index=o365 sourcetype="o365:management:activity" Operation="Consent to application."
3| eval app_name=mvindex('Target{}.ID', 0)
4| stats earliest(_time) as first_seen, values(ClientIP) as src_ips, values(UserId) as users by app_name
5| where first_seen > relative_time(now(), "-24h")
6| sort - first_seen

Detect potential JWT replay from multiple source IPs within a single token’s validity window:

1-- Splunk SPL: JWT reuse across multiple IPs (adjust index/sourcetype for your IdP logs)
2index=auth sourcetype="idp:access" token_type="jwt"
3| stats dc(src_ip) as unique_ips, values(src_ip) as ips, earliest(_time) as first, latest(_time) as last by token_id user
4| where unique_ips > 1 AND (last - first) < 3600
5| sort - unique_ips

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.