
24 Hours to Report an Exploit Your Model Can't Log
Learn how to meet the 24-hour reporting rules for embedded AI. Master edge provenance, model reconstruction, and CSIRT notifications.
Starting 11 September 2026, manufacturers selling connected products into the EU face a strict 24-hour deadline to report actively exploited vulnerabilities to ENISA and national CSIRTs under Article 14 of the Cyber Resilience Act (Regulation (EU) 2024/2847).
This reporting requirement applies immediately to products already fielded, including gateways deployed years ago. While traditional software components generate standard stack traces and memory dumps during a security incident, embedded machine learning models leave no traditional failure logs. Most CRA compliance frameworks overlook model-level failure reconstruction entirely.
The 24-Hour Reporting Sequence
Article 14 takes effect over a year before the broader CE-marking and essential cybersecurity requirements land on 11 December 2027. To remain compliant, engineering teams must execute a three-stage response sequence the moment an active exploit is detected:
The reporting clock starts the moment your team achieves reasonable certainty that an active exploit exists. Waiting for full forensic confirmation before filing violates Article 14 rules, exposing organizations to penalties up to €15 million or 2.5% of global annual turnover.
Why Dependency Scanners Miss Model Vulnerabilities
Most CRA readiness programs rely heavily on Software Bill of Materials (SBOM) generation and automated dependency scanning. While effective for identifying CVEs in deterministic libraries, static dependency scanners cannot evaluate model behavior.
The CRA explicitly identifies data poisoning and adversarial evasion attacks as reportable vulnerability classes. A dependency scanner will report a clean runtime environment even while a vision model misinterprets stop conditions due to adversarial inputs, or a predictive maintenance model outputs corrupted inferences from poisoned sensor data.
Building the Edge Evidence Pipeline
Confirming whether an inference anomaly stems from data poisoning, an adversarial input, or natural distribution drift requires capturing decision evidence at the edge before an incident occurs.
To ensure your edge system produces verifiable forensic evidence, engineer your telemetry to log three core metrics for every critical inference:
- Record the raw sensor input payload at the device boundary.
- Log the exact model version, architecture, and weights executing the inference.
- Store the resulting decision output prior to downstream execution.
Attempting to patch an edge model via an Over-the-Air (OTA) update before capturing runtime input records destroys the primary evidence needed for the 14-day final report. Telemetry preservation must precede hot-patch distribution.
Reconstruction in Practice: Drift, Attack, or Poisoning?
Here is what that evidence triplet buys you when the clock is actually running.
A predictive maintenance model on a fielded gateway starts flagging a cluster of implausible inferences. Nothing crashed, nothing threw an exception, and the dependency scan is clean. You have 72 hours to tell regulators what happened — and the answer determines whether you file at all.
With the input, model state, and output logged for every critical inference, the investigation follows three concrete steps:
Step 1: Rule out a tampered artifact. Hash the model weights running on the affected devices against your release manifest. A mismatch means a compromised update channel — a conventional security incident with its own escalation path. A match narrows the cause to the inputs.
Step 2: Replay the decision. Feed the logged raw inputs through the same pinned model version in a sandbox. If the anomalous outputs reproduce exactly, the behavior is input-driven and deterministic — you now have the exploit mechanism section of your 72-hour notification, demonstrated rather than guessed.
Step 3: Classify the input. Compare the anomalous window's input distribution against your fleet baseline:
- Gradual shift, fleet-wide, physically plausible values → natural distribution drift. An engineering problem, not a security incident. No Article 14 filing — and you have the evidence to document that determination if an auditor asks later.
- Abrupt onset, single device or small subset, structured perturbations that no physical sensor would produce → adversarial input. The 24-hour clock is running.
- Anomalies correlated with one upstream data source or a specific training window → suspected data poisoning. The clock is running, and the affected pipeline segment is identifiable in the filing.
Every branch of that triage is decidable inside the 72-hour window only because the triplet was being captured before the incident. Without the logged raw inputs there is nothing to replay, "reasonable certainty" becomes guesswork, and the choice collapses into filing late or filing a false alarm — both of which carry regulatory cost.
One Evidence Pipeline, Two EU Regulations
Under Article 12 of the CRA, products classified as high-risk AI systems under the EU AI Act that meet CRA cybersecurity obligations are deemed compliant with the AI Act's Article 15 cybersecurity requirements. Establishing a single, unified edge evidence pipeline satisfies the technical documentation standards of both frameworks simultaneously.
Preparing the Incident Response Plan
Validating CRA readiness requires testing your operational response before live enforcement:
- Set explicit metrics for distinguishing benign model drift from adversarial manipulation to prevent filing delays.
- Configure immutable local logging for inference outputs, firmware versions, and raw sensor inputs directly on edge gateways.
- Assign dedicated primary and deputy filing leads responsible for submitting notifications to ENISA's Single Reporting Platform.
System Readiness Checks
Our First Build helps engineering teams evaluate and upgrade edge architecture for regulatory compliance:
If you ship connected products into the EU with embedded machine learning models:
Submit your system architecture specs, or Schedule an architecture review with our team.
Related Insights

Insight
Moving Clinical AI from a Data Export to a Live EHR
Clinical AI models that look finished on a de-identified export routinely stall against a live EHR. The 2026 FHIR mandates standardize the pipe, not the pipeline: what still breaks is terminology mapping, stale reads, and provenance, and all three are engineering problems.
Read MoreInsight
Engineering Medical Devices from Silicon to Clinical Workflow
Moving telemetry from a medical device into a clinical workflow requires more than working hardware. This guide details the 6-layer architecture, from firmware and BLE gateways to FDA 524B compliance and RPM billing, needed to ship connected medtech into live clinical environments.
Read More
Insight
Dark Factories in Healthcare and Automotive Are a Fantasy (For Now)
In healthcare and automotive, the fully autonomous "dark factory" runs into a wall: these industries define "done" through evidence and human accountability no agent can produce for itself.
Read MoreFacing a problem like this one?
Tell us what the business needs to win and who owns it on your side. If it is a fit, a First Build starts within weeks. If it is not, we will tell you why.