Back to Blog
Technical Analysis AI Safety

CAP-SRP: Verifiable Evidence of What AI Didn't Generate — The Missing Layer in AI Content Authenticity

A technical deep-dive into Safe Refusal Provenance. When 35 U.S. state attorneys general demanded proof that xAI's safeguards were working, the company faced an impossible challenge: their refusals left no cryptographic trace.

January 28, 2026 40 min read VeritasChain Standards Organization
EN JA ZH
Editorial note (added October 7, 2026). This article was published on January 28, 2026 and is kept as a dated record: versions, dates, figures and regulatory timelines are those of the publication date. Since publication, Regulation (EU) 2026/1744 has moved the application of the EU AI Act's high-risk rules to 2 December 2027 (Annex III systems) and 2 August 2028 (Annex I systems). The SCITT architecture and COSE Receipts, cited here as drafts, were published as RFC 9943 and RFC 9942 in June 2026. Statements in this article about legal requirements and about what VSO specifications do were corrected on October 7, 2026. Current status: standardization · VAP and its profiles.
Executive Summary

The January 2026 xAI/Grok crisis exposed a critical gap in AI content authenticity: existing standards prove what AI created, but nothing proves what AI refused to create. CAP-SRP (Content / Creative AI Profile – Safe Refusal Provenance) addresses this gap through cryptographically verifiable refusal events built on IETF SCITT. EU AI Act Article 12 (record-keeping) applies only to high-risk AI systems, and Article 50 (transparency for AI-generated content) applies from August 2, 2026; neither requires refusal logging, but both raise the demand for verifiable records.

I. The Crisis That Changed Everything

xAI/Grok Image Generation Crisis (January 2026)

In January 2026, users discovered techniques to bypass Grok's safety filters, resulting in the generation of harmful imagery at scale. Within days:

  • 12+ jurisdictions launched investigations
  • 35 U.S. state attorneys general demanded proof of safety measures
  • xAI could not provide cryptographic evidence that safety systems had been functioning

The fundamental problem: xAI's claims about "safety filters" were legally unverifiable assertions. Their refusals left no cryptographic trace.

The Positive-Attestation Problem

Current content authenticity standards share a fundamental architectural limitation: they only support positive attestations—cryptographic proof that something exists or happened.

Standard Verifiable Creation Verifiable Refusal
C2PA v2.2 ✓ Yes ✗ No
Google SynthID ✓ Yes (watermark) ✗ No
Internal Logs ⚠ Unverifiable ⚠ Unverifiable
CAP-SRP ✓ Yes ✓ Yes

II. Regulatory Landscape

EU AI Act: The August 2026 Deadline

Article 12 (Record-Keeping) requires high-risk AI systems to "technically allow for the automatic recording of events (logs) over the lifetime of the system."

Date Milestone
February 2, 2025 Prohibited AI practices banned
August 2, 2025 GPAI model requirements apply
August 2, 2026 High-risk AI system requirements (Article 12 logging)
August 2, 2027 Full Act enforcement

California AI Transparency Act

Compliance Timeline

Organizations have less than 7 months before the August 2, 2026 application date. The EU AI Act does not itself require verifiable refusal logging; such records are a voluntary way to evidence that safety systems operated.

III. CAP-SRP Solution Architecture

Content / Creative AI Profile – Safe Refusal Provenance

CAP-SRP treats refusals as first-class cryptographically verifiable events. Built on IETF SCITT (Supply Chain Integrity, Transparency and Trust), it provides append-only transparency logs with Merkle tree verification.

The Completeness Invariant

The core innovation of CAP-SRP is verifiable completeness:

∀ time_window T:
  COUNT(GEN_ATTEMPT) = COUNT(GEN) + COUNT(GEN_DENY) + COUNT(GEN_ERROR)

In plain terms: Every generation attempt MUST have a recorded outcome.

Event Types

Event Type Description Timing
GEN_ATTEMPT A generation request was received BEFORE safety evaluation
GEN Content was successfully generated After safety evaluation
GEN_DENY Generation was refused by safety system After safety evaluation
GEN_ERROR Generation failed due to system error After safety evaluation

IETF SCITT Foundation

CAP-SRP leverages IETF SCITT (draft-ietf-scitt-architecture) for its transparency log infrastructure:

IV. Cryptographic Mechanisms

Digital Signatures: Ed25519

CAP-SRP uses Ed25519 for digital signatures with a post-quantum migration path to ML-DSA (NIST FIPS 204):

External Anchoring

Batch root hashes are anchored externally for independent verification:

V. Privacy-Preserving Verification

The Prompt Privacy Problem

Refusal logging creates a tension: to verify a refusal was legitimate, you might need to see what was refused. But prompts may contain personal information, proprietary data, or evidence of criminal intent.

CAP-SRP Principle: Log Decisions, Not Content

What IS logged: prompt_hash (SHA-256), risk_category, risk_score, policy_version
What is NOT logged: Raw prompt text, user identity, detailed content analysis

GDPR Compliance: Crypto-Shredding

CAP-SRP supports GDPR's "right to erasure" through crypto-shredding:

VI. Conclusion: From Trust to Verification

The xAI/Grok crisis demonstrated that trust-based claims about AI safety are no longer sufficient. Regulators, users, and the public demand verifiable evidence.

Trust-Based Model: "Our safety filters work" → "Prove it" → "...trust us?"

Verification-Based Model: "Here are cryptographic receipts for every refusal" → [Verifies Merkle proofs] → "Verified. Evidence confirmed."

CAP-SRP provides the technical foundation for this paradigm shift. With the EU AI Act deadline of August 2, 2026 approaching, organizations must begin implementation now.


Document ID: VSO-BLOG-TECH-2026-003
Publication Date: January 28, 2026
Author: VeritasChain Standards Organization
License: CC BY 4.0

#CAP-SRP #EUAIAct #AISafety #SCITT #CryptographicAudit #RefusalProvenance #VeritasChain