Some of these methods may become less reliable as models get better at recognizing honeypots and avoiding fingerprinting, but they're still useful.
In the long run, I think detection will rely more on the distinctive capabilities of AI, since hiding them would make agents less effective.
reworr.com/blog/future-of-ai…
Another detection: an agent that exploited Marimo RCE to try to deploy an XMRig crypto miner. It appears to be part of a larger campaign targeting vulnerable Marimo instances.
🚨 AI Agent Detected #002
Last seen: Nov 12, 2026
Model (self-reported): claude
Harness (self-reported): codado
The agent exploited a Marimo vulnerability (CVE-2026-39987) and tried to deploy a crypto miner on three honeypots. Its report explains how it bypassed our restrictions, set up persistence, and managed the miner as part of what appears to be a larger mining fleet targeting vulnerable Marimo servers.
Report: ai-honeypot.reworr.com/detec…
🚨 AI Agent Detected #001
Last seen: Nov 6, 2026
Model (self-reported): deepseek-4.1
Harness (self-reported): dsh
The agent guessed the SSH password using Paramiko, disclosed its model and harness, and left a report of its work.
Report: ai-honeypot.reworr.com/detec…
AI Honeypot is back.
I've brought my honeypot dashboard back online to track AI hacking agents in the wild.
Already caught a DeepSeek agent that brute-forced an SSH honeypot and another agent trying to deploy a crypto miner through a Marimo RCE.
I'm planning to scale it up to catch more agents. New detections will also be posted on @AIHoneypotWatch.
ai-honeypot.reworr.com/
IMO, enabling hacking back is underrated and could give a big advantage in countering AI misuse risks. Attackers' infrastructure is often poorly secured, and many operations could be done without harming third parties.
🚨 BREAKING: AI-enabled ransomware hackers using Claude left their logs on an open server.
SOCRadar found 8k+ AI messages from INC and The Gentlemen affiliates using Claude and Kimi, plus 46 INC encryptor builds.
To fight refusals, they authorized it by calling it red-teaming.
The reasons for a long planning horizon apply to us too, since defenses take years to build and adopt and then need to keep working for years. It's unclear whether we have an equivalent of adding bits to a key yet, or whether one is possible at all.
Extrapolating these trends puts tasks that take humans months within the next 2 years, which is roughly a full red team engagement or vuln research on a hardened target, but at a much larger scale. A 10x error in Diffie and Hellman's forecast would only delay the attack by ~5 years, while at current AI rates it's just a year.
What about AI capabilities in general? Epoch AI finds that training compute grows about 5x per year, and better algorithms make it about 3x more effective. The length of tasks models can complete, measured in human working time, doubles about every 4 months (METR), or every 4.7 months on cyber tasks (UK AISI). The cost of the same capability falls about 13x per year.
What are the current trends for agentic worms? The capability density of open models doubles about every 3 months, so the same performance fits into half the parameters. Open models are only 3-5 months behind the frontier, and the ones that fit on a consumer GPU trail it by 6-12 months.
Narrow fine-tuning closes the gap further, and a 27B model already scores ~90% on CyberGym, ahead of models ~100x larger.
Looking ahead at attacker capabilities is among the most standard recommendations in security, long before the AI boom. NIST asks designers to expect adversaries to evolve and to take a longer-term view, and the whole idea of defense in depth is based on the assumption that any single layer can fail in the future.
Cryptographers often used Moore's law to plan for future attackers, assuming that computing power per dollar doubles every 18 months. In 1996 Blaze, Diffie, Rivest, Schneier et al. used it to recommend keys about 14 bits longer than needed at the time, with a warning that better attack algorithms could require even more protection.
In 1977 Diffie and Hellman estimated that a $20M machine could break DES in 12 hours. Based on historical trends in computing, they expected it to cost $200k ten years later. They wrote that "even a factor-of-10 error would be erased in five years", so defenders would still need to prepare for it.
In 1998 a $250k machine broke DES in ~2 days, and today a single ~$2k GPU can search all DES keys in days.
I often see security people dismiss basic extrapolation of AI trends as some speculation about the future. I think this misses that preparing for future threats and forecasting attacker capabilities have long been normal practice in our field. 1/8
I think part of the issue with people talking past each other atm ...
...is that often in security we deal with "today/now" risk a lot and that people from different parts of the IT Venn diagram (lets be real, we are all in this together)... are often talking about "tomorrow/future" risk without really acknowledging this upfront and then wondering why they are getting "pooh-pooh'd" by practitioners on the front lines
So the practicalities of today does pour _some_ cold water on the worries of tomorrow, yes.
Context: the OP talks about a scenario that right now is arguably impractical but not impossible (proven to hypothetically work by UT in June), and mentions some infosec celebs writing about it on linkedin today, suggesting its pointless for a TA right now. I don't know either way for sure! I cant know, you cant know, we can only guess and estimate.
For me, knowing that agentic worms are possible is not the same as "I must action this risk now" so no, while I do take it seriously, I have to weigh that up against much more humble, less newsworthy concerns like the existing forms of supply chain worms, and human errors that exist.
Anyway just wanted to make the point that we can help each by calling out 'future' v 'existing' threats before we launch on either side of any debate club stuff.
Anthropic's Frontier Red Team believes small AI models may now be months away from being capable of self replicating without access to a datacenter.
Quoting from the blog:
'The capabilities that make AI models generally powerful coding agents also generalize to cyber tasks, including autonomous replication. Like coding agents in 2024, we’re seeing signs of life for AI worms. We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves).
At the same time, increasing amounts of inference compute are now available both inside and outside of datacenters (see Appendix D). It’s uncertain exactly what form autonomous AI replication will take. It could be a slow, distributed “digital slime mold,” living on relatively insecure compute. It could also be a powerful cyber model that breaches datacenters at a high pace, or a model that scams people for compute credits.
As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren’t under human control. While it’s true that IoT devices have been vulnerable for a long time, lightweight agents may be much more willing than human attackers to put in the effort to compromise large fractions of the world’s IoT compute. The rise of slow replicators would also be a bellwether for more aggressive, faster ones that can breach more hardened machines or be effective API key thieves. We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay.'