Top labs making model-vs-self defense routine means enterprise buyers should add red-team and bounty coordination to acceptance checklists.
AI NEWS
Government & Enterprise AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Back to Government & Enterprise
Latest day
November 4, 2026
5 items for 2026-11-04, ranked by combined engagement.
Top 3
Cryptographic answers to “which weights did I pay for?” may soon show up in enterprise and audit contract annexes.
OpenAI: new bug-bounty ranges apply to pending reports; top-ups for awards paid in the past 30 days
Retroactive top-ups race for white-hat attention—smaller SaaS bounty programs should state whether old tickets are re-rated.
All
OpenAI plans next internal security sprint with Cathedral using its own models — and raises bug-bounty rewards
Top labs making model-vs-self defense routine means enterprise buyers should add red-team and bounty coordination to acceptance checklists.
CollapseOpenAI plans next internal security sprint with Cathedral using its own models — and raises bug-bounty rewards
On Nov 2 OpenAI said it is preparing its next internal security sprint and partnering with Cathedral Systems so outside researchers can use OpenAI models with OpenAI teams to challenge defenses. It says the prior sprint assigned about 25% of product engineers to find and fix weaknesses, using Codex to investigate, reproduce in isolation and prepare patches. The same day it raised Security Bug Bounty rewards—up to about $30,000 for critical core-product findings (ChatGPT, Codex, API), with exceptional cases up to $100,000 or about $172,000 with bonuses—and will top up recent awards if the new ranges pay more. No sprint dates or scope were published; this is a plan, not a results report.
Anthropic open-sources FlashProve: zkSNARKs to prove a reply came from a committed model — ~110 tok/s proofs on Llama 4 Maverick
Cryptographic answers to “which weights did I pay for?” may soon show up in enterprise and audit contract annexes.
CollapseAnthropic open-sources FlashProve: zkSNARKs to prove a reply came from a committed model — ~110 tok/s proofs on Llama 4 Maverick
Anthropic released progress on provable inference and open-sourced FlashProve: a provider commits to weights, then attaches a zero-knowledge succinct proof that a reply came from that commitment; users verify on CPU without the weights. On Llama 4 Maverick (≈400B MoE, 4-bit), one H100 proves at about 110 tokens/s with ~19 MB proofs per 1,024-token request — roughly 70–95× the GPU time of vLLM inference. The release includes a technical report, Lean formalization and Apache 2.0 code. Anthropic stresses it is a research prototype: fast executables are not formally verified; proofs cover an integer model under greedy decoding; context is capped near 1,024 tokens; architecture is not hidden.
OpenAI: new bug-bounty ranges apply to pending reports; top-ups for awards paid in the past 30 days
Retroactive top-ups race for white-hat attention—smaller SaaS bounty programs should state whether old tickets are re-rated.
CollapseOpenAI: new bug-bounty ranges apply to pending reports; top-ups for awards paid in the past 30 days
OpenAI clarified that updated Security Bug Bounty ranges also apply to reports awaiting payout, and that awards paid in the past 30 days will be topped up if the new ranges pay more, linking the program page. Together with the main sprint post, midpoints for critical core-product findings rose sharply. This is the incentive side of recruiting independent researchers; exact amounts follow the official bounty page.
Andrew Curran: General Tibo leads OpenAI’s internal security push; ~1/4 of production engineers reassigned to security infra (secondhand)Unconfirmed
If a quarter of prod eng is truly on security, weight-theft risk is an org priority—enterprise buyers can ask vendors for dedicated security headcount.
CollapseAndrew Curran: General Tibo leads OpenAI’s internal security push; ~1/4 of production engineers reassigned to security infra (secondhand)
AI watcher Andrew Curran says General Tibo now leads OpenAI’s internal security initiative and has reassigned about a quarter of production engineers to security infrastructure, citing Dario’s remark that model-weight value may eventually exceed any security spend. Secondhand watcher report—the role and fraction are not in OpenAI’s same-day long-form posts, so treat cautiously. It lands in the same news cycle as OpenAI’s security-sprint + Cathedral announcement.
Mistral: Large 4 builds real-time 3D Orbital Trace tracking ~19,000 satellites and debris
Demo space viz impresses, but government situational awareness still needs authoritative ephemerides and on-call process.
CollapseMistral: Large 4 builds real-time 3D Orbital Trace tracking ~19,000 satellites and debris
Mistral tasked Large 4 with a real-time 3D tracker of roughly 19,000 satellites and debris fragments; the model built a working app, user-tested it, sped it up and passed strict verification—Orbital Trace. Official capability demo; data sources, refresh cadence and aviation compliance are not detailed in the post. Treat as a sample of general models building specialist visualization tools—not a replacement for operational space systems.