What a breach notice should say when the attacker is automated

A disclosure that names the attacker's method as "offensive agentic techniques" tells readers how the intrusion moved, but not where it started or how long it ran. That is the position Chainalysis put itself in on October 30, when it published a security incident notice and the full text of the email it sent to affected organizations. The company says it detected "an external actor using offensive agentic techniques to gain unauthorized access to Chainalysis systems and internal tools."

I am not going to guess at how that happened. Chainalysis has not said, and BleepingComputer reports that the company has not shared which vulnerabilities were exploited, which internal tools were accessed, or how long the attackers had access. What I want to look at is the notice itself, because more security teams are going to write one like it, and an automated attacker changes what readers need from it.

What Chainalysis disclosed

The notice follows a structure most incident responders will recognize: what happened, data impact, what the company has done, what users can do, and what happens next.

On impact, Chainalysis says it found no indication of unauthorized access to investigation content such as graphs, notes, or cases, to transaction or query results, to passwords in clear text, to MFA secrets, or to user session tokens. It says some non-sensitive data was accessed, specifically user IDs, full names, email addresses, and organization role labels for users provisioned on its platform. It is still working out which records were accessed for each organization and promises a detailed breakdown when that analysis is done.

On response, it says it engaged third-party cybersecurity experts and is using large language models to support a broader review of its platform's security, work that "extends beyond the exploited vulnerabilities." On user action, it asks people to watch for unexpected password reset requests and authentication prompts, and to turn on two-factor authentication. BleepingComputer adds that users were also warned about scammers impersonating Chainalysis.

The entry point and the number of affected users are not in the notice. Those are honest gaps at this stage of an investigation, and I would rather see them left blank than filled with a guess.

Why "agentic" changes the reader's questions

BleepingComputer describes "agentic" as AI systems that run multi-step tasks with limited human intervention, which could let an attacker automate parts of an intrusion instead of working each step by hand. Whatever the specifics turn out to be in this case, the label sets expectations for the people reading the notice.

When readers hear that an intrusion was automated, three questions follow quickly. How fast did it move from first access to the data it touched? How many systems did it try, and how many did it reach? And did detection keep pace, or did the defenders find out after the automated part was over? A human-paced intrusion lets readers assume some time between steps. An automated one removes that assumption, so the timeline carries more weight than it used to.

The notice says the activity happened "this month" and that some recipients had been notified earlier. That is enough for a first update. It is not enough for the follow-up.

What I would put in a disclosure about an automated intrusion

None of this requires revealing anything that helps the next attacker. It is about giving affected organizations what they need to judge their own exposure.

• A timeline in hours, not weeks: when the activity started (as best you know), when it was detected, when it was contained, and when you notified. Mark each one as confirmed or estimated.
• The method stated at the level you can support. "Agentic" is a claim about how the attacker worked. Say what evidence supports it, such as request rate, tool patterns, or volume of attempts, without publishing the technique.
• Scope expressed as what was touched and what was ruled out. Chainalysis already lists specific data types it found no indication of access to. The list of what was reached should be just as specific, and so should the evidence behind each exclusion.
• What changed in detection and response, in addition to what was patched. If the attacker moved at machine speed, readers want to know whether your alerting and containment can now move at the same speed.
• The next update date. "We will provide updates as our investigation progresses" is standard. A date is better.

The one place I would push hardest is the gap between "ruled out" and "found no indication." Those are different statements. The first means you looked and have evidence it did not happen. The second can mean you have not finished looking. Readers deserve to know which one they are reading.

What this means for your own incident plan

Most incident response plans assume a human attacker working at a human pace. The communication templates assume the same. If your plan says legal and communications get a draft notice within a set number of days, ask whether the facts that notice depends on can be gathered faster when the intrusion itself took hours.

Three things are worth checking before you need them:

• Logging retention and resolution. An automated intrusion can generate many actions in a short window. If your logs are sampled or summarized, you may not be able to reconstruct the sequence later.
• Pre-written impact statements by data type. The Chainalysis notice names data by type. You can only do that quickly if you already know what data lives in which system. Build that map now, while nothing is on fire.
• A rule for wording. Decide in advance who signs off on terms like "ruled out," "no evidence," and "agentic," and what proof each needs.

I wrote about keeping monitoring out of an agent's reach in the agent should never hold write access to its own monitor. That post was about your own agents. This one is the mirror case: someone else's agent working against you, and your monitoring being the record of what it did.

The questions to rehearse now

Chainalysis says it will share further updates as its investigation progresses. I will read them with interest, mostly for the timeline.

In the meantime, the useful exercise is internal. Take your current breach notice template and assume the attacker was automated. Can you fill in the timeline in hours? Can you list the data types you can rule out with evidence, separately from the ones you only have no indication about? If either answer is no, that is the work to schedule before the notice has to be written for real.