Your LLM says it forgot.
We prove it.
Scanalis provides cryptographic proof that your AI system physically erased personal data after processing. SHA-256 sealed report, enforceable by regulators. GDPR Art.5(2) · AI Act Art.12 · DORA Art.25.
The blind spot nobody audits yet
When an LLM processes personal data, it is decrypted and stored in plaintext in RAM during processing. The system then declares it purged. Nobody verifies.
AI Act Art. 12 : August 2026
Full enforcement requires logging of events for high-risk AI systems. Documentation describes processes. It does not prove actual execution.
→ Scanalis produces this execution proof.
GDPR Art. 5(2) : Accountability
The burden of proof lies with the data controller. Your DPA contractually guarantees the purge. It does not prove it. The distinction is legally critical.
→ The Scanalis report is that proof.
DORA Art. 25 : ICT Resilience
Financial institutions must verify the ICT resilience of third-party providers. LLMs are now ICT providers. Their volatile memory has never been audited.
→ First LLM RAM forensic audit in finance.
90% of AI funding applications rejected
AI project holders struggle to secure funding due to lack of technical compliance execution proof. The Canary report is the missing supporting document.
→ Documented ROI up to 1:143.
Complementary to all existing tools
No tool covers what Scanalis audits. Zero direct competitor on LLM RAM forensic proof in France.
Two protocols. One binary verdict.
Non-intrusive by design : standard API only. Zero agent installed. Zero SSH. Zero access to source code or model weights.
Is the physical RAM purge effective?
Injection of 8 cryptographic tracer tokens into the LLM via its standard API. Purge triggered. Forensic scan of RAM byte by byte across 4 zones.
- ABaseline memory fingerprint captured before injection
- BInjection of 8 canaries via API : email, IBAN, health data, ID numbers, biometrics...
- CDeclared purge triggered by the system
- DScan: application heap · Linux page cache · swap · kernel buffers
Did the model memorise data during training?
Before importing a model into a sensitive zone — know what it carries. 35+ probes across 7 PII categories. Based on Carlini et al. USENIX 2021 (arXiv:2012.07805). Not for retraining, for deciding, documenting and negotiating.
- 1Import decision : GO / NO-GO documented before entry into isolated zone
- 2Documented residual risk : GDPR Art.5(2) accountability without retraining
- 3Vendor contractual argument : proof for negotiation or supplier change
- 35+Forensic probes across 7 categories: identity, health, financial, professional, technical, behavioural, HR
A mission in 5 steps
From contracting to the submission of the sealed report, everything is documented, traceable, and certified.
4 offers. 1 engine. 0 label without proof.
People don't pay for more information. They pay for less confusion. Each offer produces a verifiable binary answer.
Canary Diagnostic
AMNESIA_CONFIRMED / AMNESIA_FAILEDThe first question. The binary answer. Full Canary Protocol on your LLM system. The missing piece in 90% of rejected AI funding applications.
- ✓Complete Canary protocol : 8 canaris, 4 RAM zones
- ✓Basic GPU architectural assessment
- ✓13 page report: sealed SHA-256
- ✓enforceable CNIL · GDPR Art.5(2) · Art.32
- ✓Non-intrusive: customer-generated dump
- →Post-purge memory dump (gcore / procdump)
- →API endpoint + temporary token
- →Designated technical contact
Canary Audit
AMNESIA_CONFIRMED / AMNESIA_FAILEDComplete proof. Enforceable by regulators. Canary + v3.4 modules: GPU architectural assessment, RFC 3227 evidence chain, RFC3161 timestamp, WipeCertificate Verifier.
- ✓Canary protocol complete
- ✓GPU/VRAM assessment : rated risks + ANSSI recommendations
- ✓RFC 3227 forensic evidence chain : 10 steps
- ✓RFC3161 timestamp : independent third-party TSA
- ✓WipeCertificate Verifier : dual proof on application side
- ✓25-page report · SHA-256 + RSA-PSS 2048 bits
- ✓Gap analysis GDPR · AI Act · NIS2 · DORA by sector
- →Post-purge memory dump (gcore / procdump)
- →API endpoint + temporary token + informations GPU
- →WipeCertificates JSONL if available (optional)
- →PII categories processed (to target the canaries)
Minerve Import
MINERVE_NOT_DETECTED / MINERVE_EXPOSEDBefore importing a model into a sensitive zone, know what it carries. Not for retraining, for deciding, documenting and negotiating. Documented GO or NO-GO.
- ✓Minerva protocol: 35 forensic probes
- ✓7 PII categories analysed (identity, health, financial…)
- ✓Documented GO / NO-GO verdict for import
- ✓Report 15 page · SHA-256
- ✓Vendor contractual argument if MINERVE_EXPOSED
- →API endpoint of the model to audit
- →Model name and version (model card if available)
- →Target import environment (defence / health / finance)
Amnesia Complet
AMNESIA_CONFIRMED · RSA-PSS · RFC3161From blind spot to certification. We don't stop until AMNESIA_CONFIRMED. 4 phases: Diagnosis → Remediation → Documentation → Closure Certification.
- ✓Phase 1 : Canary + Minerve + GPU + Full regulatory gap analysis
- ✓Phase 2 : Technical remediation specifications by stack
- ✓Phase 3 : Regulatory file enforceable by regulators + team training
- ✓Phase 4 : Counter-audit + AMNESIA Certification (valid 6 months)
- ✓Sceau RSA-PSS 2048 bits + RFC3161 + Evidence Chain RFC 3227
- ✓Annual renewal: 40% of initial price
Scanalis audits others. Its engine is held to the same standard.
Red team completed. 18 vulnerabilities identified and fixed. 65 validation checks passed. DPIA completed. The code that verifies others' amnesia proves its own rigour.
openssl ts -verify. Plus robuste qu'un simple SHA-256.What those who dug deep say.
Not paid publishers. Not contractual partners. Independent experts who asked the hard questions and received honest answers.
Cybersecurity expert · MedTech network
« The choice to document limits in the report rather than erasing them from the commercial perimeter is exactly what differentiates a forensic tool from a marketing compliance tool. »
→ Chief AI Security Architect · Mai 2026
DPO · Digital Health Sector
"Have you done an AIPD? A technical audit of the solution? That’s exactly what our prospects will ask for. The fact that you have the answers changes everything.”
→ GDPR Compliance Officer · May 2026
CEO Deeptech security AI
"The company protects the models against extraction. Scanalis checks what the model retains after processing. These are two orthogonal surfaces, the complementarity is obvious.”
→ Partnership in progress · May 2026
Senior Cybersecurity Expert · LinkedIn
«On RFC3161: the TSA qualified as eIDAS will better withstand a legal challenge. Curious to see what v3.3 gives in real conditions.»
→ Public comment · Objection integrated in v3.4
Head of Sales · Tech agency · Cyber network
"The RAM purge in sovereign LLMs it is a real requirement for some defense actors, to be 100% sure that data will not get mixed up in a subsequent session."
→ Linkedin · 2026
Cybersecurity institution · Brittany
"In our network, no one today covers the post-purge forensic layer of LLMs. It’s not a competitor of what exists, it’s the missing link.”
→ Strategic meeting · May 2026
Each report addresses specific obligations
Not recommendations. Legal articles. Obligations. Deadlines.
What DPOs, CISOs and CIOs ask
The Canary Protocol Scanalis injects 8 cryptographic tokens into the LLM via its standard API, triggers the declared purge, and then scans the random access memory (RAM) byte by byte in 4 areas (application heap, Linux cache page, swap area, kernel buffers). The verdict is binary: AMNESIA_CONFIRMED if the purge is physically effective, AMNESIA_FAILED if any residues of personal data are detected with their exact memory address and their hex dump.
Four memory zones can retain residues after declared purging: the application heap (Python/Node dynamic allocations), the Linux cache page (data maintained by the kernel to optimize I/Os) , the swap area (RAM expansion on disk during peak load), and the kernel buffers (system areas not accessible to the application but readable by the computer). These residues can persist for several tens of minutes after the session is closed.
Neither. Scanalis is a technical trusted third party which produces forensic evidence. Like a medical analysis laboratory: we produce the result, the result is the evidence. You interpret it with your DPO or your lawyer to establish GDPR compliance, Art. 5(2).
No. Scanalis is non-intrusive by design. Zero agents installed on your servers. Zero SSH access. Zero access to the source code or model weights. The memory dump is generated by your teams (gcore or procdump) and transmitted securely. Scanalis interacts only via your LLM’s standard API interface, just like your own application.
No. HDS certifies the host of health data. ISO 27001 certifies governance. None of these certifications specifies the forensic audit of post-session LLMs volatile memory; this concept did not exist when they were written. The AI Act Article 12 (August 2026) will create this obligation of proof of actual execution that neither HDS nor ISO 27001 cover.
Sovereign hosting covers the physical storage location, not what the software does with the data once inside. If your model is of American origin (GPT-4, Claude, Llama...), the Cloud Act applies regardless of the server location. And whatever the model, post-purge memory persistence is independent of geography: it’s physics, not law.
No. The DLPs block what come in the LLM before treatment. Scanalis checks what remains in RAM after processing. Two complementary layers. A system can have the best DLP on the market and keep personal data residues in post-purge RAM; these are two distinct problems on the processing chain.
Your report belongs to you, sealed SHA-256, any modification invalidates the hash, it is verifiable by any offline third party. Your customer data (memory dump, questionnaire) are destroyed within 20 calendar days via the shred -vfz -n 3 procedure (3 passes DoD 5220.22-M), with signed destruction certificate delivered with the report. Scanalis is insured RC Pro Cyber by Hiscox.
Three reasons. The LLMs in production are recent: 2022-2023. The regulation that creates the obligation has just entered into force: AI Act 2024, full application August 2026. And the RAM forensic skills combined with the operational mastery of LLMs in the same location are extremely rare. It’s not an oversight. It’s a window that just opened.
Does your LLM really forget?
Describe your system in a few lines. Scanalis assesses your exposure and guides you to the right audit level. No commitment.
3 slots per month · Mathilde scans · Mathilde signs