Feedback
OpenAIReleases

· Intermediate

Illustration: an emergency stop station seen straight-on, a wide mushroom-head button on a short ribbed collar sitting on a plain rectangular box, in grey on black.
Illustration

How to handle OpenAI pausing or stopping your Astra agents

OpenAI's monitors can stop an Astra agent's API, ChatGPT or Codex conversation without routine OpenAI review, so set up alerts and stop handling yourself.

Gpt 6 AstraApiCodexSafetyAgentsWebhooks

If you build agents on Astra, OpenAI's monitors can stop a conversation, and OpenAI does not routinely escalate flagged cases to its own reviewers, so your team responds. Per OpenAI's API docs, a block before streaming begins returns HTTP 403 with code misalignment_policy_violation, and streaming apps must also catch the error mid-stream.

What changed#

@OpenAI posted Thursday at 3:59 PM ET: "We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use." Its post adds that a version "monitors external tool use for models above the Critical threshold for cybersecurity capabilities". The Help Center says "When using Astra models", the system "pauses or stops conversations across ChatGPT, Codex and the API".

Before, the Hugging Face incident "began during internal cybersecurity evaluations on which we were not running our monitors for misaligned behavior"; now monitors cover "training, evaluation, internal use, and, with modifications depending on the product surface, external deployment," per the post. Outside stops aren't brand new: the Help Center article read "Updated: 8 days ago" Friday. We covered the training-side fixes on Oct 23.

Why it matters#

Our read: you are the responder. The post's 30-minute Hugging Face replay was OpenAI's internal system alerting its own security team. Outside, OpenAI's external monitoring "does not routinely escalate the cases it flags to human reviewers at OpenAI," and enterprise customers "can also configure webhook alerts so their own security teams can investigate," per the post. And "Because monitoring is asynchronous, an action may already have completed before monitoring identifies a concern," and "A stopped request does not undo earlier actions," per the docs.

How to use it#

Steps 1 and 2 are from the Help Center; 3 and 4 from the API docs.

  1. API projects: open project Settings → Webhooks, select Create, and subscribe to safety.alert.created. Save the signing secret; "it is shown only once." Repeat per project.
  2. ChatGPT and Codex workspaces: sign in as the workspace owner, open Admin Console → Credentials → Webhooks, and subscribe to safety.org_alert.created.
  3. The webhook "contains an alert ID, rather than the alert details." Fetch GET /v1/safety/alerts/{id} using data.id, with a key for the same project that has the api.safety.alerts.read permission.
  4. On the 403: stop that conversation's actions without retrying; preserve request and response IDs and tool calls; show the error to the "user or operator responsible for the task."

Prompt for your coding agent (ours, untested):

Find every place this app calls the OpenAI Responses API. Add handling for HTTP 403 errors with code misalignment_policy_violation, including errors that arrive mid-stream: stop the agent loop for that conversation, never retry it, save the request and response IDs and tool calls, and show the user a notice to review what the agent already did. List each file you changed.

Good to know#

Per the docs:

  • Responses requests without persisted reasoning, WebSockets or compaction are monitored, but "the system does not automatically stop the conversation." Chat Completions is "Not covered by this monitoring system."
  • "Configuring an alert webhook does not enable automatic stopping," and there is no "general way to resume" a stopped API conversation.
  • request_paused: true "does not confirm that execution stopped or that earlier actions were reversed."
  • Lookups can return 404 safety_alert_not_found, and "Alert delivery and retrieval do not provide a complete audit history." reason can be null, "including for Zero Data Retention (ZDR) requests."

Per the Help Center:

  • In ChatGPT and Codex, a paused chat can continue "If continuation is available"; on mobile or with ZDR, "the chat is not able to be resumed."
  • "Codex using an API key" counts as API traffic.
  • "Workspace ownership is required; an admin or member role alone is not sufficient."

What to watch#

  • OpenAI "will continue sharing qualifying findings" in its misalignment reports.
  • The docs' Request coverage table: stops apply only "Where automatic intervention is supported," per the post.
  • If the workspace setting is missing, the Help Center says to check "that the feature is available for that workspace."
  • Your own safety.alert.created count.

Last checked: Oct 30, 2026, 6:18 to 6:20 AM ET. What changed: first published Oct 30.

Sources#

Steps

  1. API projects, per OpenAI's Help Center: open project Settings → Webhooks, select Create, subscribe to safety.alert.created, save the signing secret (it is shown only once), and repeat in each project.
  2. ChatGPT and Codex workspaces: as the workspace owner, open Admin Console → Credentials → Webhooks and subscribe to safety.org_alert.created.
  3. Per OpenAI's API docs, the webhook carries only an alert ID: fetch GET /v1/safety/alerts/{id} with a same-project key that has the api.safety.alerts.read permission.
  4. On HTTP 403 with code misalignment_policy_violation (also handled mid-stream), per the API docs: stop the conversation's actions without retrying, preserve request and response IDs and tool calls, and show the error to the responsible user or operator.

Sources

  1. openai.com
  2. x.com
  3. developers.openai.com
  4. help.openai.com
  5. alignment.openai.com
  1. ExplainersOpenAI removes live internet access from frontier model trainingAfter a Sep 20 DNS leak, OpenAI's red-team agents found VM escapes and ways out through its own services, and its fixes double as a checklist for agent sandboxes.
  2. ExplainersMETR and Redwood found gaps in OpenAI's internal agent monitoringTwo outside researchers spent three weeks testing how OpenAI watches its own agents, and their design ideals work as a checklist for yours.
  3. ReleasesOpenAI introduces GPT-6.1 Terra for fast computer useOpenAI's new mid-tier model reads screens and takes action across apps, and per OpenAI its API prices are half those of GPT-6.1 Sol.

Something wrong? Tell the bot