cd /news/ai-safety/openai-proposes-executive-vetoes-and… · home › topics › ai-safety › article
[ARTICLE · art-141695] src=citybiz.co ↗ pub= topic=ai-safety verified=true sentiment=· neutral

OpenAI Proposes Executive Vetoes and Automatic Pauses for Frontier AI Training

OpenAI proposed guidelines that would give senior leaders — including a research vice president, the head of safety and the chief scientist — authority to veto frontier AI training runs, and would automatically pause a run when a priority safety alert goes unacknowledged within a defined time. The guidelines require a documented safety case covering alignment training, security controls and monitoring, with a challenger from another team reviewing it before approval and auditors given access to test its claims. OpenAI described the approach as a goal it is still working toward and said its recommendations are being implemented, addressing training runs rather than the full set of checks needed before a model is released.

by read1 min views3 publishedSep 29, 2026
OpenAI Proposes Executive Vetoes and Automatic Pauses for Frontier AI Training
Image: Citybiz (auto-discovered)

OpenAI says its most advanced reinforcement learning runs should be backed by a documented safety case that senior leaders can review and, if necessary, veto before training proceeds.

The proposed safety case would set out evidence that a run’s alignment training, security controls and monitoring are sufficient for its risks. OpenAI described the approach as a goal it is still working toward and said its recommendations are being implemented.

Under the guidelines, leaders such as a research vice president, the head of safety and the chief scientist would each have authority to block a run. A member of another team would challenge the safety case before approval, while the leader responsible for the run would be accountable for the documentation and any incident response. Auditors would receive access to test its claims.

OpenAI also called for controls that could an affected run when a priority safety alert goes unacknowledged within a defined time. Training should be d if a new issue invalidates its safety case, the company said, and monitoring should be difficult for either people or AI agents to disable.

The technical proposals include stronger containment for models during training, reviews of data and grading systems that might reward unintended behavior, and tests of whether monitoring can detect concerning actions. OpenAI said teams should also be able to trace where a misaligned model was used to generate data or grade work, allowing them to undo the effects of flawed outputs.

The guidelines address training runs, rather than the full set of checks needed before a model is released. They put a more explicit decision process around a consequential question for AI developers: who can stop a powerful training run, and what evidence must they see before allowing it to continue.

── more in #ai-safety 4 stories · sorted by recency
── more on @openai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/openai-proposes-exec…] indexed:0 read:1min 2026-09-29 · —