The Log Is Also the Breach

Read the Warp disclosure closely and the uncomfortable part is not the percentage. BleepingComputer reported on Oct 27 that agent decision logs and telemetry belonging to users of Warp, an agentic development environment with at least 800,000 users, were exposed in an attack on a third-party vendor in Warp’s infrastructure. Founder and CEO Zach Lloyd estimated that roughly 1 percent of users may have been impacted, said the operation was detected within hours, and said CrowdStrike was engaged and law enforcement notified. Then the sentence that matters: the company did not clarify whether the exposed information included source code, prompts, credentials, or other sensitive details.
That gap is the story. A vendor held the records. The records held agent traces. And at disclosure time, the company could not say in plain terms what was inside its own traces. Most enterprises are in the same position, they just have not been asked the question in public yet.
This is the second incident of this shape in a week. The same report notes that OpenCode notified users last week that a breach exposed stored API keys used with external providers, affecting about 1.4 percent of users and leading to unauthorized charges of up to $1,700 for some. I wrote about that incident in Decommission the Keys First, from the retirement angle. Warp adds the other half. First the keys a tool stores. Now the logs a tool keeps. The supporting data around an agent, the data nobody classifies, is becoming the payload.
An agent trace is not ordinary telemetry. Telemetry says a feature was used, a call took 400 milliseconds, a process crashed. A decision log says what the agent was asked, what it planned, which tools it called with which arguments, which files it read, and what it returned. String a week of those together and you can reconstruct prompts, project structure, internal names, snippets of source code, and the reasoning a developer gave the machine. If a secret passed through the agent’s context, and secrets do, it can sit in the trace in plain text. That is a work record and a source archive in one file, and most shops store it with the controls they use for crash reports.
The irony is that we keep these records for good reasons. I argued in Publish Your Own Incident Log that a shop running agents needs its own record of what its agents did, written before an investigator or a regulator asks. I still hold that. A record you cannot produce is a governance failure. But a record you produce, ship to vendors, retain forever, and leave readable by half the company is a different failure wearing the first one’s clothes. Governance created the collection. Almost nobody finished the job by classifying what was collected.
The fix is to split the record into two layers that get different treatment. The first layer is the audit summary: the agent identity, the owner, the action taken, the target system, the outcome, the timestamp, and a hash of the full trace. That layer is small, structured, and safe to keep for years, because years is how long accountability questions take to arrive. The second layer is the full content trace: prompts, tool arguments, outputs, files touched. That layer is sensitive, short-lived, encrypted, and access-controlled, with a retention period measured in weeks unless an incident holds it open. When the full trace expires, the summary remains, and the hash proves the summary was not rewritten after the fact. You keep the ability to answer what happened long after you stop keeping the material an attacker would want.
Four tests show whether a shop has done this work or only talked about it.
First test: sample your own traces. Pull a random day of agent logs from one team and read them the way an attacker would. Look for credentials, tokens, customer records, personal data, and code that was never meant to leave the repository. Do the reading yourself, not with a scanner report summary. Every shop I have seen do this finds something. The finding is not the scandal. Never having looked is.
Second test: map where traces go. The Warp attack did not hit Warp’s core product. It hit a vendor in its infrastructure. Ask each tool on your estate where decision logs and telemetry are stored, which subprocessors touch them, what region they sit in, how long they are kept, and whether they are used for training or product improvement. Get the answer per data class, not per marketing page. If a vendor cannot separate traces from telemetry in its answer, that is your answer about how it stores them.
Third test: run the one-hour scope question. If the vendor holding your traces disclosed a breach tonight, could you state within an hour which classes of your data were in those traces for the affected period? Warp could not clarify prompts, code, or credentials at disclosure. That is not a criticism of one company’s incident response. It is the normal state of shops that never inventoried trace content. The time to build the inventory is before the call, because during the incident the inventory is the difference between a scoped notification and a vague one that forces you to assume the worst.
Fourth test: check who can read traces inside your own walls. Trace access tends to spread quietly: support, product, analytics, engineering, the vendor success team. Each group has a plausible reason. Together they mean the most complete record of what your developers and agents did together is one of the most widely readable datasets in the company. Restrict it like source code, log access to it, and review that access on the same cycle you use for production systems. I made a related point in The Monitor Must Live Outside the Agent’s Reach: the oversight record needs its own protection. Protection includes confidentiality, not just tamper resistance.
None of this argues for keeping less of a record. It argues for keeping the record on purpose. Log what accountability requires. Store the summary like an audit record and the content like the sensitive material it is. Know which vendors hold each layer, and be able to say what is in your traces before someone else reads them first.
Thinx-Tank