Chainalysis breach shows the *evidence gap* when AI agents attack

What happened at Chainalysis

In October, Chainalysis detected external attackers using what the company called 'offensive agentic techniques' to breach their systems and internal tools ^[1]. The blockchain analytics firm said there was no indication that investigation data had been unauthorised accessed, and that the incident was contained ^[1].

The phrasing is careful, and it has to be. When attackers use AI agents that adapt their approach in real time, proving what they did and didn't reach is harder than proving that something was detected.

Think of it like a burglary where the intruder keeps changing tools. You can show the broken lock and the footprints, but showing which rooms they entered and which drawers they opened is a different question.

Why agentic attacks are different

Offensive agentic techniques means using AI agents to probe defences, find paths through systems and adjust tactics as they go ^[2]. Traditional attacks follow a script. Agentic attacks write the script as they run, in response to what they find.

That makes them harder to detect in the moment, and harder to reconstruct afterwards. The logs that tell you what happened are written by the same systems that were under attack, often while the incident was still unfolding.

The evidence gap

Chainalysis is in the business of tracing blockchain transactions, so it knows a thing or two about evidence ^[1]. Still, saying there is 'no indication' that data was accessed is not the same as proving it wasn't.

The gap is this: logs are kept by the people being attacked. If an attacker had administrator access during the breach, they could, in theory, edit the logs on their way out. That doesn't mean they did. It means the logs alone can't rule it out.

For a court, an insurer or a regulator, that gap matters. A statement that there is no indication of data access is an honest summary of what the logs show. It isn't proof that the logs are complete.

What the law expects

Cyber resilience rules, like those under DORA in finance or the Swedish Cybersecurity Act, don't just ask organisations to report incidents ^[4][5]. They expect you to be able to show what happened: which systems were affected, when you knew, and what you did.

The Swedish law, for example, says organisations must send an early warning within 24 hours of becoming aware of a significant incident, then an incident report within 72 hours ^[4]. The clock starts when you knew, not when it was convenient.

If the logs that tell you when you knew can be quietly edited, the report is harder to trust.

Sealing logs during an incident

This is where a digital witness fits. The thing to seal is each log segment as it closes, and the first detection note as soon as the incident is spotted. Sealing takes a fingerprint of the log, not the log itself, so the data never leaves your systems.

If a question arises later, you can show that the log is exactly what was written at that time. Change a single character in it and the fingerprint won't match.

What it doesn't do is show that the log is complete, or that the response was adequate. It just makes it harder for someone to rewrite the timeline after the fact, and easier for you to show that you didn't.

You seal the records through WitniumChain, using the REST API, the SDK or the MCP server. The content stays with you. Anyone you share a log with can check the fingerprint themselves. The first 2,000 witnesses a month are free, and you can read the technical details in the documentation.

How it would work

1. Your monitoring system detects unusual activity and writes a detection note. That note is sealed immediately.
2. As the incident unfolds, log segments from affected systems are sealed every few minutes.
3. When you report the incident to the regulator, you can show that the logs you're relying on haven't changed since they were written, even by someone with administrator rights.

What this means for you

If you run systems that could be targeted by adaptive attacks, ask yourself:

• Could we show, six months from now, when we first detected an incident and what the logs said at the time?
• If an attacker had administrator access during a breach, would that create doubt about whether the logs are complete?
• When we report an incident, can we back up the timeline with records that nobody could have quietly edited?

The hardest part of an agentic attack isn't detecting it. It's proving afterwards what happened, to a standard that satisfies an auditor, an insurer or a court.

What to watch

The EU's Cyber Resilience Act applies from 11 September 2026, with detailed reporting requirements for manufacturers of connected products ^[6]. AI agents are already being used in both attack and defence, and the logs that reconstruct an incident are only as trustworthy as the systems that keep them.