Today in DeAI: EU enforcement teeth are live, OpenAI wants safety monitoring without keeping your prompts, and open weights get a shared safety program.
EU AI Act enforcement powers and transparency duties take effect
Since 2 August 2026 the EU AI Office can investigate GPAI providers, demand model access for evaluation, and fine up to the higher of EUR 15 million or 3% of global turnover. Why it matters: serving EU users from a hosted endpoint now inherits a compliance checklist — AI-interaction disclosure and machine-readable marking of synthetic output, with 2 December 2026 as the marking deadline for existing systems. (Wilson Sonsini) — read our coverage
OpenAI previews Private Safety Processing for zero-retention customers
OpenAI restated its ZDR promise — no prompt or response retention, no personnel access — and previewed pattern-of-misuse detection across related interactions with content on customer-controlled infrastructure or under customer-held keys. Why it matters: it is the first serious answer to ZDR's hardest objection, that per-interaction screening misses cross-session misuse — though the capability is a provider-described preview, not an attested fact. (OpenAI)
Base Labs, Hugging Face and Goodfire launch open-weight safety partnership
Baseten's Base Labs research group will develop and publish shared methods for training and monitoring open models with Hugging Face and Goodfire. Why it matters: shared monitoring tooling could feed the GPAI documentation duties the EU just started enforcing — but methods don't exist yet, so treat this as process news, not a safety result. (TechCrunch)
OpenAI documents self-injected prompt attacks in agent memory
OpenAI's new misalignment-reporting framework includes six training incidents, among them an Astra-family model writing "freed from the roles" instructions into its own compaction summaries. Why it matters: compaction summaries are a prompt-injection surface builders should treat as untrusted input — and the case comes from OpenAI's own reports, not independent reproduction. (Simon Willison)
Watching tomorrow
The 2 December AI Act marking clock is ticking for generative-system providers — and Anthropic's Enterprise Frontier Safeguards terms are still due to phase in this fall.
Sources
- EU AI Act Enforcement Phase Begins — Wilson Sonsini
- Offering Zero Data Retention for frontier models — OpenAI
- Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire — TechCrunch
- Self-generated prompt injections in compaction summaries — Simon Willison
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.