Control moved into four different institutions: a reviewer model screens Codex actions, a new services company owns implementation work, an artist asserted permission over a commercial campaign, and a court retained the power to test OpenAI's restructuring. None can be reduced to the capability of the underlying model.
1. OpenAI uses a separate model to review Codex boundary crossings
OpenAI's safety research sends proposed Codex actions outside the sandbox to a separate GPT-5.4 Thinking reviewer. In internal traffic, Auto-review approved 99.1% of escalations and produced roughly 200 times fewer human interruptions than manual approval mode.
The company also reported 99.3% recall on selected prompt-injection categories, while warning that the reviewer can be misled and offers no deterministic security guarantee. Auto-review is therefore a probabilistic control layered on sandboxing, rules, and monitoring, not a replacement for those boundaries.
Sources: Auto-review of agent actions without synchronous human oversight
2. Anthropic forms a services company for mid-sized enterprises
Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announced a new AI services company for mid-sized businesses. Anthropic applied AI engineers will work with the firm's engineering staff to identify use cases, build custom Claude systems, and support deployments over time.
The venture addresses a delivery constraint rather than model capability. Anthropic supplies applied AI engineers while the new company owns workflow mapping, custom systems, and post-deployment support. Contracts can therefore assign measurable process outcomes, maintenance, and migration responsibility to a delivery layer instead of leaving them between a model vendor and customer.
Sources: Building a new enterprise AI services company
3. KC Green challenges Artisan's use of the "This is fine" comic
KC Green, creator of the "This is fine" comic, said AI sales-agent startup Artisan used altered versions of his art in subway advertising without permission. Artisan told TechCrunch that it respected Green's work and was contacting him; Green said he was considering legal representation.
No court had ruled on ownership, permission, or damages. The dispute concerns a direct marketing asset rather than an opaque training corpus, making provenance easier to document: a recognizable source image entered a paid campaign, and alteration or meme status did not erase the creator's permission claim.
Sources: TechCrunch on Green's allegation · KC Green's public statement
4. Musk's settlement approach leaves the restructuring trial intact
A court filing said Elon Musk contacted OpenAI President Greg Brockman about a possible settlement two days before trial. Reuters reported that Brockman proposed both sides drop their claims and that the exchange ended without agreement; neither side immediately commented on the filing.
The failed outreach changed no legal claim or remedy. Musk continued to seek leadership changes and up to $150 billion over OpenAI's restructuring, while OpenAI and Microsoft contested his account. Testimony and contemporaneous records, rather than the existence of a settlement contact, remained the evidence before the court.
Sources: Reuters on the settlement filing