Signal Watchtower Advisory SW-A006: AI Reconnaissance, a Year After the First Warning

The Signal

Preemptive security is now Gartner’s recommendation

I have been warning about AI-driven reconnaissance for more than a year. I first wrote about it on 31 October 2025, before Anthropic disclosed the incident known as GTG-1002 in November 2025.

Gartner’s latest Cybersecurity Executive Outlook reports that 76% of CISOs rank AI-driven vulnerability discovery among their top 10 emerging risks. More importantly, Gartner recommends investment in preemptive cybersecurity, specifically exposure management, predictive threat intelligence and deception.

This validates my concerns for the past year because these are issues I have been writing about and researching. Here is the timeline of my writing.

01The Record

Twelve months of writing, in order

On 31 October 2025, I published an Elytra Signal Watchtower Critical Advisory introducing the hypothesis of Systematic Automated Reconnaissance (SAR). My concern was that increasingly automated, potentially AI-assisted reconnaissance could systematically discover exposed infrastructure, misconfigurations and weak trust relationships at a speed defenders were unprepared for at that time. Sadly, that applies even today.

Our defenders are still too slow to match the attackers’ speed.

It was a hypothesis then, and I was convinced that this would not be the first incident. I stand vindicated now, after observing so many more such incidents in the last year.

ETR³ 2025 (Elytra Threat, Risk & Resilience Report) warned that we would begin to see a lot of such activity in 2026. Part IV said mass reconnaissance was already routinely generated or assisted by large language models, compressing the time from targeting to exploitation, and its predictions for 2026 listed AI-assisted reconnaissance among the factors that would materially raise breach likelihood. As I said in a recent LinkedIn post, I predicted that SAR would become a significant challenge for enterprise cybersecurity over the following 12 to 18 months. I was wrong, and overly optimistic. Anthropic’s GTG-1002 disclosure was released in November 2025. It documented AI-orchestrated cyber espionage, and attributed it to Chinese nation-state actors.

I continued writing about SAR, examining the emerging evidence. In my December 2025 follow-up I published a defensive playbook and argued that enterprises needed to adopt systematic reconnaissance themselves. My research also led me to build Argus and Venus at Elytra Security.

The regular editions carried the line forward. Edition 006 (30 April 2026) was the first regular edition after the SAR Critical Priority Advisory. It argued that trusted systems left over-permissioned or weakly monitored become the terrain on which accelerated reconnaissance operates. Edition 007 (15 May 2026) extended it when Google reported what it assessed as the first known AI-assisted zero-day exploit. I wrote then that the danger is that AI is beginning to remove the discovery bottleneck.

02The Playbook

Reconnaissance on your own estate

This is the SAR playbook as it appeared in ETR³ 2025, Part IV. It has three steps.

Assume attackers have tireless interns. They are called models. Find the issues yourself before they do.

The Sequence We Work To

In Honey, I Shrunk the Defence (Executive Signals, August 2026) I set out the sequence our own work follows. It carries the same playbook into daily operations.

When OpenAI published its research on the Hugging Face incident, which I examined in Advisory SW-A002, its timeline recorded an agent recovering 14 publicly exposed Hugging Face credentials with write access. Leaked keys are on the playbook’s list. I made the same point in The Keys We Leave Outside: credentials that remain valid and reachable can turn one point of access into several.

03What We Built

Argus for exposure, Venus for behaviour

Argus and Venus are Elytra Security products.

What Venus has recorded so far

So far, Venus has recorded over 10 million external events from more than 54,000 source IP addresses. We are using this data to study reconnaissance patterns, including those potentially consistent with agentic behaviour.

Source: Venus dashboard, latest batch processed 12 November 2026 at 04:03 UTC. The dashboard shows 10.1 million external events from 54.5 thousand distinct external sources, out of 14.6 million events recorded across all sources.

I would not say that we can confidently attribute anything to AI activity. But we have some markers that do lead us to suspect as such. We have seen about 1,160 events that are suspected to be AI reconnaissance. A paltry ratio, no doubt, but this is just one deception platform. Imagine the events that might be on your own systems right now.

04Why It Resonates

Where Gartner and our evidence meet

And this is why Gartner’s recommendations resonate with me. We have spent a year developing the arguments, gathering evidence and engineering defensive capabilities around them. I have maintained throughout that security teams need to understand their own exposure at least as systematically as their adversaries attempt to discover it.

This is the most important credo for CISOs now:

When reconnaissance operates at machine speed, exposure management cannot remain a periodic exercise.

Questions For Your Team

In my 30 October article I proposed measures for this. They deserve the same management attention as vulnerability counts and incident statistics.

05The Evidence Trail

For anyone interested in how this thinking developed, the original publications and subsequent research are listed below.

1. SAR Critical Advisory, 31 October 2025
2. ETR³ 2025, Elytra Threat, Risk & Resilience Report
3. SAR follow-up, evidence and defensive playbook, December 2025
4. Signal Watchtower Edition 006, 30 April 2026
5. Signal Watchtower Edition 007, 15 May 2026
6. Signal Watchtower Advisory SW-A002: The Machine Found Another Way, 30 July 2026
7. The Wild Signals: The Keys We Leave Outside, 3 August 2026
8. Executive Signals: Honey, I Shrunk the Defence, 26 August 2026
9. Venus operational research, 29 October 2026
10. Argus and Venus, 30 October 2026
11. Venkat Mangudi on LinkedIn: SAR prediction and the wake-up call
12. Gartner: Cybersecurity Executive Outlook, 28 October 2026
13. Gartner: Top Five Actions for CISOs to Take by End of 2026, 30 September 2026
14. OpenAI: The Hugging Face incident and the road ahead