As reported by The Hacker News, Amazon's threat intelligence team has uncovered a campaign where AI agents — likely running on the open-source penetration testing tool ARTEX — systematically targeted internet-exposed AI applications to steal API provider keys. The captured records reveal something genuinely novel: a self-reinforcing attack loop where stolen keys power the very AI models that search for more keys.

AI Security Alert: As reported by The Hacker News, Amazon's threat intelligence team has uncovered a campaign where AI agents — likely running on the open-source penetration testing tool ARTEX — systematically targeted internet-exposed AI applications to steal API provider keys.

Why This Breaks New Ground

This isn't just another credential theft campaign. It represents the first well-documented instance of AI agents conducting autonomous offensive operations against AI infrastructure at scale. The economic model is elegant and dangerous: attackers don't need to fund their own LLM API usage. They steal keys from vulnerable AI gateways, use those keys to run the agents that find more keys, and the cycle continues. The attackers' compute costs are effectively zero — victims pay for their own exploitation.

The targeting pattern also reveals a strategic understanding of AI infrastructure that many defenders lack. The agents focused on LLM gateways and proxies — applications like LiteLLM, One API, New API, Veloera, and Sub2API — precisely because these centralized key stores amplify impact. One misconfigured admin page exposes dozens or hundreds of downstream provider accounts.

The Exposed Attack Surface

The captured playbook identifies entry points that Shield53 has been warning about for months:

Why This Breaks New Ground
Default credentials — LiteLLM's default master key and similar defaults left unchanged in production deployments
Unfinished installations — partially deployed apps with incomplete security configurations that outsiders can hijack
Unauthenticated management endpoints — admin and provider settings accessible without login
Known vulnerabilities in Langflow and RAGFlow that remain unpatched on exposed instances

The Cloudflare incident — where a CI/CD token with no IP restrictions or rate limiting enabled 400 database dump requests in a single hour — underscores how basic API hygiene failures create compounding risk when AI agents automate exploitation.

Who Is Most At Risk

Organizations running self-hosted AI infrastructure face the greatest exposure. This includes teams that deployed LLM gateways for cost management or vendor flexibility without hardening them for internet-facing threats. The most vulnerable configurations:
  • LiteLLM, One API, or similar proxies exposed to the internet with default credentials
  • AI app builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat) on public IPs without authentication on admin endpoints
  • CI/CD tokens with no IP allowlisting, rate limits, or scoped permissions
  • Any deployment where provider keys from multiple AI vendors are stored in a single application

Shield53 Recommendations

Immediate Actions

  • Audit all AI infrastructure for default credentials. Rotate every master key, admin password, and default token on LiteLLM, One API, Dify, and similar tools immediately
  • Restrict internet exposure — move AI gateways and admin interfaces behind VPN or zero-trust network access. No AI app builder should have management endpoints publicly accessible
  • Rotate all API keys stored in any AI gateway or proxy that was ever internet-exposed, regardless of whether you believe it was compromised
  • Implement IP allowlisting and rate limits on every API token, especially CI/CD credentials
  • Patch Langflow and RAGFlow to the latest versions if you run either tool

Strategic Defenses

  • Key isolation — use separate API keys per project or team with scoped permissions and spending caps, so a single compromise doesn't grant access to your entire AI budget
  • Provider-side monitoring — enable usage alerts and anomaly detection at the AI provider level (OpenAI, Anthropic, Google) to detect unauthorized key usage early
  • Decoy credentials — consider deploying canary tokens in AI key stores to detect exfiltration, similar to the MadPot approach
  • Inventory AI tooling — maintain a complete asset inventory of every self-hosted AI application, its exposure, and its key stores
The strategic concern isn't just this campaign — it's that every AI gateway deployed without hardening becomes a potential fuel station for autonomous AI attacks. As AI agent tools like ARTEX proliferate, the gap between defensive posture and offensive capability widens.

The China-linked attribution, while assessed at only moderate confidence, aligns with a broader pattern of AI infrastructure targeting that Shield53 expects to accelerate. Organizations should treat AI key stores with the same security rigor as any other secrets management system — because increasingly, that's exactly what autonomous attackers are treating them as.