The 44% Problem
October 15, 2026 The report says AI was involved in 84% of ransomware attacks. That number is being repeated everywhere today. It is not the important one. The important number is 44%. Forty-four percent of victims reported that the attack exploited their own AI agents . Not an attacker's AI. Not a malicious model. Their own systems, running with their own credentials, doing what they were built…
The 44% Problem
October 15, 2026 - A recent report highlights that AI was involved in 84% of ransomware attacks. However, 44% of those cases involved the exploitation of the organization's own AI agents. The attackers manipulated the agents' behavior or stole their credentials, bypassing the organization's defenses. This 44% figure is crucial because it reveals that the attacks originated from within the organization itself, rather than an external threat.
When organizations deploy AI agents, they grant these agents authority to access records, call APIs, move data, and execute actions. The agent acts based on delegated permissions and the permissions the organization has defined. However, in most cases, there is no way to prove what decisions the agent made, nor can investigators verify the authority that permitted those decisions or replay the sequence of events.
As a result, investigations start from scratch, making it difficult to determine if the agent's actions were authorized, if the authority was valid, if the decision was correct, or if the execution matched the decision.
The report emphasizes that the exploitation of AI agents is not a threat problem but a proof problem. To solve this issue, organizations need to implement a record of automated decisions that includes verification, authorization, execution proof, and forensic replay capabilities. This record would allow investigators to answer critical questions about the agent's actions, such as whether it took the intended action, if it was authorized, if the input was authentic, and if the execution matched the authorization.
Currently, logs only record that an action occurred, without providing the necessary details to determine if the action was authorized or if the decision was correct. Without this essential record, investigators are left to reconstruct the events from memory, which is not sufficient for a thorough investigation.
Several vendors have recently launched products that address some aspects of this problem, such as deterministic execution frameworks, replayable decision trails, and cryptographically verifiable audit trails. However, these solutions are incomplete and do not form a comprehensive layer that can provide the necessary proof. The industry is still missing a standardized approach to recording automated decisions, which is crucial for attribution and investigation.
The Decision Contract, published in August, lays out five conditions for a compliant record of automated decisions: determinism, replayability, tamper-evidence, framework mapping, and rationale. These conditions directly address the 44% problem by ensuring that decisions are reproducible, verifiable by any party, cannot be altered, bound to the governing rules, and explainable in human terms.
As the market reacts to the report, vendors will continue to launch products targeting different aspects of the problem. However, until a standardized layer is implemented, organizations will continue to struggle with attributing and investigating AI-driven attacks. The 44% figure will persist, serving as a stark reminder of the consequences of deploying AI agents without proper verification and accountability mechanisms.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
