Since June we have been running Venus, our deception platform, with deliberately constructed services exposed to the internet. Venus records who turns up and what they do.
Traffic arrived on the first day.
In the first 24 hours Venus recorded 22,500 external events. They came from more than 1,100 IP addresses across 184 distinct networks and touched the web, database, shell, portal and API surfaces.
As of 29 October 2026, Venus has recorded about 7.7 million external events from more than 45,000 distinct source IP addresses. The dashboard also holds roughly 11.8 million events across all feeds, which is a different scope and is not used in this article. In the last seven days Venus recorded about 768,000 external events. Over the last 30 days our analytics identified 19,160 source entities, of which 6,376 returned. Of those, 378 showed activity classified as progressing beyond initial reconnaissance, about 2%.
What Venus is
Venus is a deception platform. It exposes services that have no legitimate users, so every connection to them is unsolicited. There is no business traffic to filter out, and what remains is a record of what the internet does when it finds something.
Most security telemetry is collected from systems that serve real people, and the hostile activity has to be picked out of the legitimate. Venus starts from the other end. Because the traffic is all unsolicited, we can study behaviour in its own right: how sources return, how sessions connect, how activity changes between services.
For defence, this means evidence from our own environment instead of borrowed statistics. It also gives us a place to test questions that production systems cannot answer safely, including how automated and AI-assisted tooling behaves when it meets an environment it does not know.
What are they looking for?
Most visitors run routine scans. Others probe for exposed credentials, administrative interfaces, vulnerable applications and services that should never have been public. We see repeated attempts against SSH, RDP, SMB, databases, web applications and authentication interfaces.
A source that returns over several days, changes its probing or moves between services gives us far more behavioural evidence than one that sends thousands of requests in a few seconds. Venus therefore looks at sessions, persistence, recurrence and sequence alongside event volume.
Persistence
Our latest 30-day analysis found 346 highly persistent sources. Some returned across several weeks and generated hundreds of sessions. Misconfigured automation and scheduled scanners produce the same pattern, so persistence alone says nothing about sophistication.
It does let us study how sessions relate to each other, and whether activity moves from discovery to credential access attempts. The 378 multi-stage classifications narrow the work to sequences that merit a closer look.
AI cadence
We have started to look for AI-assisted reconnaissance through behaviour. Venus carries three separate indicators, each shown as a badge. A known AI bot is an identification tag. An AI-associated egress marks a source address linked to AI service infrastructure. A potential SAR is our own heuristic for patterns consistent with Systemic AI-assisted Reconnaissance.
None of the three establishes that an autonomous agent generated the traffic. An address on AI infrastructure may carry any request, and an automated sequence can resemble agentic behaviour with no model involved.
So we examine cadence: timing between requests, repeat sessions, movement between surfaces, and changes in the order of requests. A conventional scanner follows a predetermined list. An autonomous agent could change its next action based on what a response told it. The open question is whether any observed sequence shows decisions made on information discovered during the visit. We are working on that question.
Deception and autonomous attackers
Recent research on Context Bombs extends conventional deception. Misleading context is placed where an autonomous agent will read it, in the hope of sending the agent after false leads or making it spend resources. Honeypots, honey credentials and decoy infrastructure have been used for years. What is new is the chance to watch how an agent responds over successive interactions.
Could a believable decoy credential change an agent’s next action? Would it drop one line of investigation for another? Could we separate a fixed workflow from one that adapts? These are research questions. Venus telemetry does not show that Context Bombs have redirected an autonomous agent in our environment. It does give us a large body of behavioural evidence and the analytical tooling for controlled tests.
Limits of the evidence
An IP address can front several clients, and one operator can use many addresses.
A known bot or AI-associated egress address is useful intelligence, but it does not identify the software or purpose behind a given request. Behavioural scores and cadence analysis are leads for investigation and need validation before they support any claim about autonomous AI activity.
The observations also reflect the services, configuration and exposure of the Venus environment. They are not a measurement of malicious internet activity in general. Geolocation, ASN and hosting labels describe network registration, not where an operator sits.
Research note: all operational figures come from Venus dashboard snapshots dated 29 October 2026. Behavioural and technique classifications are platform assessments. AI-related indicators and potential SAR detections are not independently verified attribution to autonomous AI agents.