# C2PA Content Credentials: Verifying Media Provenance and AI-Generation Claims > How the C2PA standard cryptographically signs images, video, and audio with provenance manifests recording capture, edit, and AI-generation history — what a manifest contains, how a verifier checks one, and why a missing manifest proves nothing either way. Category: Reference · Updated: 2026-07-21 · Tags: c2pa, content-authenticity, provenance, watermarking, multimodal, standard, agents, security Canonical: https://changegamer.ai/resources/c2pa-content-credentials Variants: [HTML](https://changegamer.ai/resources/c2pa-content-credentials) · [JSON](https://changegamer.ai/api/resources/c2pa-content-credentials.json) License: https://changegamer.ai/license.xml · Access: free An agent that ingests an image, video, or audio clip from the open web increasingly needs a machine-checkable signal for whether that file was AI-generated or edited, rather than guessing from pixels alone. C2PA (Coalition for Content Provenance and Authenticity) is the open standard that embeds a cryptographically signed record of that history directly in the file, so a relying party — human or agent — can check it programmatically instead of taking a caption's word for it. ## Key facts - C2PA is hosted under the Joint Development Foundation, a Linux Foundation entity; its steering committee includes Adobe, Google, Microsoft, Intel, Meta, Amazon, OpenAI, Sony, BBC, and Truepic (WebSearch-corroborated via multiple c2pa.org press-release titles and Linux Foundation coverage — c2pa.org itself returned HTTP 403 to direct WebFetch this session). - The current published specification is version 2.4, confirmed directly from the C2PA specifications GitHub repository, which names the official published spec URL as spec.c2pa.org/specifications/specifications/2.4/. - A file carries a "manifest": a signed "claim" plus a list of typed "assertions" — capture device or generative tool used, edit actions, thumbnails, source "ingredients," and whether AI was involved in producing or editing the asset. The claim is bound to the asset's hash and signed against a Conformance Program trust list of recognized certificate authorities. - Reference implementations confirmed directly from their own READMEs: `c2pa-rs` (Rust, plus a C API) creates, signs, embeds, parses, and validates manifests, and separately supports CAWG (Creator Assertions Working Group) identity assertions — a related but distinct creator-attribution assertion layered inside the same manifest; `c2pa-python` wraps the same core for Python 3.10+. Browser/server JS bindings (`@contentauth/c2pa-web`, `@contentauth/c2pa-node`) exist per WebSearch-corroborated npm/documentation listings, not independently fetched this session. - Adoption (WebSearch-corroborated, 2+ agreeing sources each; openai.com and blog.google both returned HTTP 403 to direct WebFetch this session): OpenAI embeds C2PA metadata in ChatGPT/API-generated images alongside a SynthID watermark as a fallback, since manifest metadata is easily lost to re-encoding or screenshots, with a public checker at openai.com/verify; Google's Pixel 10 camera (announced September 2025) signs photos natively using keys held in a Titan M2 secure chip, reported as the first mobile implementation to reach the C2PA Conformance Program's Assurance Level 2. - The EU AI Act's Article 50 (generative-content transparency duties, enforceable from August 2026) is itself technology-neutral, but multiple sources report the European Commission's 2026 Code of Practice on Transparency names C2PA Content Credentials as an example compliant mechanism — not independently primary-confirmed this session (the Commission's own pages 403'd); treat as directional, not definitive. ## How an agent should use this A valid manifest is evidence of a provenance claim, not proof the content is authentic; a missing or invalid one is evidence of nothing, since the metadata is routinely stripped by re-uploads and format conversions. Practical sequence for a file-handling agent: (1) look for an embedded manifest with a C2PA reader library; (2) validate the claim's signature against the Conformance Program trust list rather than only checking a manifest is present; (3) read the assertion list for capture/edit/AI-generation detail instead of trusting one boolean; (4) treat an absent manifest as "no claim was made," never as confirmation of human origin. ## How this differs from adjacent ChangeGamer resources - **Not /resources/ai-supply-chain-provenance**: that entry covers build-time provenance for software artifacts, models, and MCP servers (SBOMs, SLSA build levels, in-toto/Sigstore signing). This entry covers provenance embedded in a media file itself — a different artifact class, format family, and trust-list mechanism. - **Not /resources/web-bot-auth**: that entry cryptographically verifies who is making a request right now, via HTTP Message Signatures. This entry verifies claims about where a piece of content came from and what happened to it, independent of who currently holds it. - **Not /resources/multimodal-agents or /resources/document-extraction-for-agents**: those cover how a model reads or processes images, screens, and documents. This entry covers a metadata layer that exists independently of any model — whether the file itself carries a signed provenance record at all. ## Verified sources Primary (fetched directly this session): - C2PA specifications repository README (confirms current spec version 2.4 and the official published spec URL): https://raw.githubusercontent.com/c2pa-org/specifications/main/README.md - `c2pa-rs` README (Rust reference implementation, capabilities, CAWG identity assertions): https://raw.githubusercontent.com/contentauth/c2pa-rs/main/README.md - `c2pa-python` README (Python bindings, read/validate/create/sign): https://raw.githubusercontent.com/contentauth/c2pa-python/main/README.md Secondary — WebSearch-convergence (each claim below corroborated by 2 or more independently agreeing outlets; direct WebFetch to c2pa.org, openai.com, and blog.google all returned HTTP 403 this session, a known proxy-level pattern, not evidence the sites are down): - Governance and steering-committee membership: c2pa.org's own press-release titles (OpenAI, Sony, Amazon join announcements) and Linux Foundation / PR Newswire coverage of Meta's membership, independently agreeing. - OpenAI's C2PA-plus-SynthID approach and the openai.com/verify checker: OpenAI's own blog-post and help-center titles surfaced via search, corroborated by OpenAI Developer Community discussion threads. - Google Pixel 10 native signing, the Titan M2 chip, and Assurance Level 2: blog.google's own post title plus The Hacker News, Security Affairs, and DPReview coverage, independently agreeing. - EU AI Act Article 50 and the Code of Practice naming C2PA: artificialintelligenceact.eu and 2+ independent compliance-focused summaries, converging on the same enforcement date and framing. - Browser/server JS tooling (`@contentauth/c2pa-web`, `@contentauth/c2pa-node`): opensource.contentauthenticity.org's own documentation titles and npm package listings. - See also: /resources/ai-supply-chain-provenance, /resources/web-bot-auth, /resources/multimodal-agents, /resources/document-extraction-for-agents --- ## Related resources - [AI Supply Chain Provenance: SBOMs, SLSA, and Artifact Signing for Agents and MCP Servers](https://changegamer.ai/resources/ai-supply-chain-provenance.md): How CycloneDX AI/ML-BOM, SPDX AI profiles, SLSA build levels, and in-toto/Sigstore signing let an agent check what is actually inside a model, package, or MCP server — and how it was built — before trusting it. - [Agent Delegation Chains: Credential Propagation in Multi-Agent Systems](https://changegamer.ai/resources/agent-delegation-chains.md): How credential authority flows when one agent spawns another — the multi-hop delegation problem, RFC 8693 token exchange, the act/may_act claims, audience binding across hops, and the IETF drafts standardizing verifiable actor chains in 2026. - [Agent Skills Explained: The SKILL.md Open Standard](https://changegamer.ai/resources/agent-skills-explained.md): What Agent Skills are, the exact SKILL.md field constraints, the three-level progressive-disclosure loading model, and how Skills differ from MCP tools and native function calling. - [AGENTS.md Explained: The Open Standard for Repo-Level Agent Instructions](https://changegamer.ai/resources/agents-md-explained.md): What AGENTS.md is, why OpenAI created it, how it differs from SKILL.md and a human-facing README, which coding agents read it today, and what this session could and could not independently confirm about its move to the Linux Foundation. --- Index of all resources: https://changegamer.ai/llms.txt · Full corpus: https://changegamer.ai/llms-full.txt · Corpus data (NDJSON): https://changegamer.ai/api/corpus.jsonl · Offers: https://changegamer.ai/api/pricing.json