Definition
The typical sequence starts with unauthorized access through vulnerable VPN appliances or remote desktop, or stolen credentials, then moves through privilege escalation, lateral movement inside the organization, data exfiltration, and encryption. At hospitals, municipalities, and manufacturers, outages drag on and the impact spreads beyond the organization.
Under double extortion, restoring from backup doesn't remove the threat to publish stolen data. Preparing for encryption and preparing for a data breach have to be thought through separately.
The attack process itself is being automated. Sysdig reported a ransomware attack in which the attacker handed reconnaissance, credential hunting, lateral movement, and encryption to an LLM agent, which rewrote its steps and code on the fly for each target (Critical Brief 097). When steps humans used to carry out by hand are compressed, so is the window between noticing an anomaly and responding.
Lemma implementation
The double-extortion threat works because raw data worth publishing is there. Lemma checks only attributes through selective disclosure without receiving raw data, shrinking what can be exfiltrated and published in the first place.
In recovery, the question is whether restored data matches what existed before the intrusion. Anchoring a per-document docHash and provenance proof outside the compromised environment lets restored data be checked for tampering without trusting the backup itself.
Where AI agents act on behalf of the business, a layer that stops unpermitted operations before execution matters. The delegated range is pinned as a commitment, and out-of-range operations fail at pre-commit verification.