DEV Community

Cover image for Attacker agents now turn patches into exploits. They still get in through your defaults
Sofia_ Humanbound for Humanbound

Posted on

Attacker agents now turn patches into exploits. They still get in through your defaults

Humanbound's engine and CLI are open source: github.com/humanbound/humanbound

On 16 November, Google's Threat Intelligence Group disclosed exploit_agent, a framework that diffs security patches, writes the exploit, and runs it with no human in the loop. Days earlier, Amazon and the FBI described ARTEX-based agents stealing keys from exposed AI gateways and using them to pay for more hunting. Both reports land on the same uncomfortable point: the agents are fast, but they get in through default credentials, open admin panels, and secrets left within reach.

What Google found

Google Threat Intelligence Group (GTIG), with Mandiant, published "Agentic threat activity crosses a significant threshold" on 16 November. The centrepiece is a multi-agent framework Google calls exploit_agent, found on an attacker staging server that was itself left exposed to the internet.

The workflow is simple to describe and uncomfortable to read:

  1. Watch web software for new patches.
  2. Diff the code before and after the fix to find what was fixed.
  3. Draft an exploit for the unpatched version.
  4. Run it automatically.

Google says this is the first time it has seen attackers fully automate vulnerability discovery and exploitation without a human involved. Follow-up reporting from CoinDesk on 18 November adds detail: the files carried Russian-language comments (translated by Google as "LLM API keys for auto-exploitation (exploit_agent pilot)"), the tool called both commercial and local LLMs, and records point at French media, entertainment, and education sites. The same server held recon scripts, brute-force tooling, and web shells.

GTIG also reports the open-source agentic frameworks CAIRN and PentAGI in real intrusions, and puts a number on the pressure: zero-days averaged 8 per month in 2025, 11 per month in 2026, and 22 in August 2026 alone.

What Amazon and the FBI found

On 13 November, AWS CISO CJ Moses and the FBI published parallel warnings about ARTEX, an open-source LLM-driven pentest framework.

Amazon's MadPot decoys, posing as LLM services, recorded attacker agents between 3 October and 5 November. They hunted for exposed AI apps, extracted stored provider keys, validated them, and pooled them to fund the next round of hunting. Amazon is careful here: it "cannot verify the outcomes the agents reported," and the ARTEX attribution is moderate confidence.

The target list reads like a typical AI platform team's stack: LiteLLM, One API, New API, Dify, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama. The entry points are even more familiar:

  • Default admin credentials, including LiteLLM's default master key
  • Unfinished installs an outsider can complete
  • Unauthenticated management and provider-settings endpoints
  • SSRF through URL-fetch features to reach cloud metadata
  • Secrets in client-side JavaScript, source maps, and debug routes
  • Known, patched flaws: Langflow CVE-2025-3248 (in CISA KEV) and RAGFlow CVE-2026-28797

The FBI's FLASH alert says ARTEX has been used since at least August, with some intrusions reaching privileged access in under three hours. Amazon's summary is the line worth pinning: "What AI agents change is the speed and persistence of the attacker." None of the techniques were new.

The pattern: speed on top of hygiene debt

Put the two reports side by side and the threat model gets clearer, not scarier:

  • The patch window is now an agent's runtime. If exploit_agent can diff a patch and ship an exploit automatically, "we'll patch in the next maintenance cycle" is a decision to stay exposed.
  • Your AI stack is now a target class of its own. Gateways and agent builders hold the most valuable secret in the building: provider keys that turn into free compute for the attacker.
  • The attacker's agent is configured badly too. Google notes these frameworks trade stealth for speed, and both exploit_agent and many ARTEX consoles (1,465 exposed hosts per Censys as of 4 November) were found because they were left open. Noisy agents leave artefacts you can detect.

What defenders shipped in response

The same days brought a visible shift toward containing agents at the platform layer:

  • Microsoft Perception went GA at Ignite on 18 November, and Agent 365 with Entra Suite (preview) can now detect 60 types of local AI agents, including Claude Code, Codex, and GitHub Copilot.
  • macOS 27.3 beta (16 November) splits device-control permissions and adds per-use prompts for automation and agent apps.
  • US Senators Cantwell and Rounds are pushing the White House AI task force toward NIST security standards for AI agents.

What to do with your own agents today

  1. Inventory every AI gateway and agent builder you run. Rotate every default credential and close half-finished installs.
  2. Take provider keys out of the agent's reach. Scoped, short-lived tokens; nothing long-lived in env files the agent can read.
  3. Patch AI tooling like production software, because attackers now diff your vendor's patches faster than you deploy them.
  4. Alert on spend and volume, not just errors. A key-harvesting agent shows up on the bill first.
  5. Attack your own agents before someone else's agent does. Run prompt injection, tool misuse, and data exfiltration tests against the agents you ship, and keep running them after every change.

Test your agents the way attackers now do. Humanbound's free Community plan gives you hosted attack models that probe your agents for prompt injection, tool misuse, and data leakage. Sign up at app.humanbound.ai.

References

Top comments (0)