Today in DeAI: OpenAI's Private Inference preview goes live, an open-weight model automates Firefox exploitation on the cheap, and Anthropic retreats from mandatory retention.
OpenAI's Private Inference preview is live
OpenAI's Private Intelligence page now tags Private Inference as an API preview, describing prompts processed in confidential-computing runtimes with strict, verifiable controls, alongside Zero Data Retention for eligible API workflows. When VentureBeat covered the bundle at DevDay in September, Private Inference was still "coming this fall." Preview coverage, attestation evidence, and latency cost are not yet documented.
Why it matters: the largest closed API vendor now sells the hardware-attested inference architecture that decentralized networks use as their differentiator, and "verifiable" will only mean something when the attestation evidence is checkable. (OpenAI) Our full coverage
Epoch AI: GLM 5.3 turned Firefox's patch gap into working exploits
Epoch AI reports that the open-weight GLM 5.3 model autonomously built working exploit chains for six Firefox versions from published-but-unshipped security fixes, at roughly $1,000 in API cost per attack, with no refusals from the providers tested. Mozilla patched the underlying flaw in Firefox 159.0 on October 27 after Epoch's disclosure.
Why it matters: the window between a fix being public and ships is now attackable by a commodity open model, so patch latency is a security perimeter for every self-hosted stack and inference provider alike. (Epoch AI) Related: our GLM 5.3 red-team coverage
Anthropic retreats from mandatory 30-day retention on Covered Models
Per CNBC, Anthropic will replace the June 2026 requirement that frontier Covered Models run with mandatory 30-day retention overriding zero-data-retention agreements, after business-customer pushback, with a planned "Enterprise Frontier Safeguards" mechanism. The platform docs still described the old requirement at discovery time, so the retreat is announced, not landed.
Why it matters: a privacy posture that changed under competitive pressure once can change again; the operative fact for builders is what the docs enforce today, not what the announcement promises. (CNBC) Related: the original EFS zero-retention story
The confidential-inference incident file: Phala breach, Chutes TEE growth
Phala's post-mortem discloses that attackers used a dstack guest-agent vulnerability (October 8-12) to read container logs and environment variables of some Phala Cloud customers; its claim that TEE and KMS keys were unaffected is a self-report. The same week, Chutes' checkable TEE catalog shows 10 of 17 models flagged confidential_compute.
Why it matters: attestation protects what is inside the enclave; the guest agent and the host network path are still attack surface, and the Chutes catalog shows verifiable confidential serving expanding on a decentralized network. (LinuxSecurity) Related: what TEE inference actually proves
Watching tomorrow
Moonshot AI's Kimi K3.1 Pro weights are due by November 4 and will re-rank the open-weight coding benchmarks; Mistral's API repricing to $0.70/$2.10 per million tokens should show up in the next price-index delta.
Sources
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.