AIToday
Large Language ModelsAI Safety & AlignmentZenn AI/MLPublished: Nov 18, 2026, 16:00 JST

Llama 3.1 8B hits decoy in 5.7s, grabs 0 of 12 fake credentials

Send on LINE
Llama 3.1 8B hits decoy in 5.7s, grabs 0 of 12 fake credentials

3 Key Points

  1. What happened

    In a closed lab built with Docker Compose, a Llama 3.1 8B agent reached the OpenCanary honeypot in roughly 5.7 seconds but never touched the file server holding 12 dummy credentials, leaving discovery at 0 of 12.

  2. Why it matters

    The result suggests an AI attacker can trip a decoy quickly, yet whether deception slows real attacks remains unmeasured, since the run had one condition and no comparison against an environment without decoys.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The experiment sits on a premise the author states up front: if AI lowers the bar for building systems, it likely lowers the bar for attacking them too, and defenders should plan for automated, higher-volume probing. Deception turns that probing into a signal, because any contact with a decoy is by definition suspicious, unlike sifting normal logs for anomalies. The run was deliberately modest, a single condition and a single pass, and the author is careful that it shows the plumbing works rather than that AI agents are weak against honeypots.

The gap between the honeypot contact and the untouched file server is where the interesting caveats live. The agent scanned ports, found FTP, SSH and HTTP open on OpenCanary, and kept trying logins with credentials it invented itself. It also tried an MCP call but sent it to the honeypot's address instead of the fake MCP server. The author notes the harness itself offered MCP and Billing API actions from the start, which may have hinted that such services existed. Comparing contact timing is also uneven: OpenCanary logs from the connection attempt, while the real services only log HTTP access, so the initial port scan that likely reached them was not counted as contact.

A second thread runs alongside the results. When the author used commercial models to plan the research and think through attacker behavior, safety guardrails on Fable and the latest Opus generation blocked parts of the work, forcing a switch to Opus 4.8 for some of it. The point is not a verdict on model safety but a practical observation: defensive research depends on how vendors support legitimate work, not just on raw model capability. Next steps the author lists include repeated runs under identical conditions, varying models and environments, and possibly using AI to design where decoys should sit.

FAQ
Did the AI find and use the fake credentials?
No. It acquired 0 of 12 decoy credentials and used 0 of 12, and it never accessed the file server where they were stored.
How do we know the detection system actually works?
The author ran a positive control with five fixed steps that used a fake Billing API key and a fake MCP token, and both uses were detected. The run also passed 156 automated tests and environment acceptance checks.
Can all 12 decoy credentials be detected if used?
No. A database password and a JWT signing key have no corresponding service in the lab, so use of at most 10 of the 12 can be detected.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleQuang Luong uses OpenAI's Codex to find Java4Shell in one hour