AI Risk Wire

DeepSeek

DeepSeek has 10 model versions on record; 2 have at least one documented fact.

Of the 2 versions listed below, 2 document tests by an independent or government party. None documents a published framework, threshold determinations, tests the provider ran itself, safeguards the provider describes.

Model versions with documented facts

ModelWhat is documented
DeepSeek-V4-Pro · 8 documents
DeepSeek-R1 · 2 documents

No document we read states a capability level, a test or a safeguard for: DeepSeek-V4.1-Flash, DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Flash-Base, DeepSeek-V4-Flash-Vision-Exp, DeepSeek-V4-Pro-Base, DeepSeek-V4-Pro-0813, DeepSeek-V4-Flash-0731. DeepSeek-V4-Flash: its documents report tests only as table rows or lines that do not name the version on their own; its page links the documents.

Each test and safeguard, by who documents it

For each item, how many of the 2 versions listed above whose documents we read in full have it from an outside party (an independent or government tester), from the provider, or as a claim we could not confirm.

No document we read states, for any of these versions: Published framework with capability thresholds; Threshold determination for this release; Serious-incident disclosure commitment; Pre-release external access; Tested: autonomy and replication; Tested: harmful content; Tested: injection and agentic risks; Tested: bias; Tested: privacy; Safety training (harmlessness and refusal); Agent permission limits and sandboxing; Human approval for high-risk actions; Misuse monitoring and enforcement; Reasoning (chain-of-thought) monitoring; Staged or restricted access; Output provenance and watermarking; Telemetry and audit logging; Domain-specific safeguards (CBRN or cyber); Vulnerability disclosure and bug-bounty programme; Third-party or government red-team access during development.