---
type: "post"
id: "preparing-our-next-security-sprint-to-harden-openai-7d9d46"
slug: "preparing-our-next-security-sprint-to-harden-openai-7d9d46"
title: "Preparing our next security sprint to harden OpenAI"
description: "OpenAI says it is preparing another internal security sprint, this time working with Cathedral Systems researchers who will use OpenAI models alongside company teams to test defenses."
retrieval_nugget: "OpenAI says it is preparing another internal security sprint, this time working with Cathedral Systems researchers who will use OpenAI models alongside company teams to test defenses. The announcement also recalls a previous effort involving 25% of product engineers and Codex. It signals a push toward continuous, model-assisted defense, but does not yet report results from the new sprint."
published_at: "2026-11-03"
updated_at: "2026-11-07"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["codex"]
entities: ["OpenAI","Cathedral Systems","Codex","Daybreak"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "same_event"
prior_refs: ["https://newruntime.com/posts/as-ai-becomes-more-capable-at-finding-vulnerabilities-1796e1/"]
source_url: "https://openai.com/index/preparing-our-next-security-sprint-to-harden-openai"
source_urls: ["https://openai.com/index/preparing-our-next-security-sprint-to-harden-openai"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:preparing-our-next-security-sprint-to-harden-openai-7d9d46"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/preparing-our-next-security-sprint-to-harden-openai-7d9d46/","markdown":"https://newruntime.com/posts/preparing-our-next-security-sprint-to-harden-openai-7d9d46.md","json":"https://newruntime.com/posts/preparing-our-next-security-sprint-to-harden-openai-7d9d46.json"}
markdown_url: "https://newruntime.com/posts/preparing-our-next-security-sprint-to-harden-openai-7d9d46.md"
---

# Preparing our next security sprint to harden OpenAI

## Retrieval answer

OpenAI says it is preparing another internal security sprint, this time working with Cathedral Systems researchers who will use OpenAI models alongside company teams to test defenses. The announcement also recalls a previous effort involving 25% of product engineers and Codex. It signals a push toward continuous, model-assisted defense, but does not yet report results from the new sprint.

## Summary
OpenAI says it is preparing another internal security sprint, this time working with Cathedral Systems researchers who will use OpenAI models alongside company teams to test defenses. The announcement also recalls a previous effort involving 25% of product engineers and Codex. It signals a push toward continuous, model-assisted defense, but does not yet report results from the new sprint.

## What happened
In a post dated November 2, 2026, OpenAI said it is planning its next internal security sprint. The company will bring together new researchers and security partners with its own teams to use its models to test and strengthen OpenAI systems. Cathedral Systems is the named partner. Its researchers will use OpenAI models alongside OpenAI teams, contributing what the company describes as differentiated expertise and approaches for finding weaknesses and helping protect its platforms. The post does not give a start date, duration, or a list of systems in scope.

OpenAI frames the planned work as part of an ongoing effort to use AI to find and fix weaknesses continuously. It says traditional cyber defenses alone are no longer sufficient, citing recent incidents across the industry in general terms. The company’s stated aim is to identify and address weaknesses before attackers do. The post offers no incident examples or threat metrics to quantify that claim.

To explain the new effort, OpenAI points to its last security sprint. It says 25% of its product engineers were dedicated to finding and fixing weaknesses in company systems. Using Codex, engineers investigated findings, reproduced vulnerabilities in isolated environments, and prepared patches for review. This is a useful description of a workflow—investigation, reproduction in isolation, then patch preparation—but the company does not report how many findings were handled, how many fixes shipped, or how long the sprint lasted.

The announcement also says OpenAI is improving its Security Bug Bounty program, including increasing rewards for critical findings by more than four times. It wants researchers to challenge its assumptions and help strengthen both its systems and defensive methods. The post links to more detail about the bounty program but gives no new reward amounts, eligibility terms, or examples of qualifying findings in this announcement.

OpenAI says it uses its most capable models to strengthen its own defenses and makes them available to Daybreak customers. It says it has committed $1 billion in subsidized access over six months for frontline defenders protecting essential services and digital infrastructure. It also says it is sharing architecture and lessons from the last sprint, alongside a blueprint for cyber defenders. These are broader parts of the company’s security program; the post does not say the Cathedral collaboration itself is part of Daybreak or specify how the partner will access models.

## What the archive adds
This is a new development on an active OpenAI security thread, not a wholly new topic. [OpenAI’s earlier account of hardening its research training environment](https://newruntime.com/posts/over-the-past-several-months-we-have-substantially-7f9243/) described layered controls and reduced attack surface. [A prior Field Note on the Defense Factory and internal security sprint](https://newruntime.com/posts/as-ai-becomes-more-capable-at-finding-vulnerabilities-1796e1/) covered the broader defensive program and the previous sprint. This announcement adds Cathedral Systems as a named external partner and reiterates the earlier sprint’s 25% engineering allocation.

## Trend signal
The announcement supports [agent security moving to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/): OpenAI describes finding weaknesses in systems and reproducing them in isolated environments before preparing fixes. The post does not describe the specific controls or establish that model use itself improves security; it is evidence of an organizational approach, not proof of its effectiveness.

## Why now
OpenAI connects the work to recent incidents across the industry and argues that defenders need to find and fix weaknesses before attackers do. It also presents this as a continuation of its prior sprint and wider effort to put frontier cyber capabilities to defensive use. The post does not identify a particular incident as the trigger for the new sprint.

## Why it matters
For teams building with AI, the most useful detail is not the claim that models can help with security, but the workflow OpenAI says it used: investigate a finding, reproduce it in an isolated environment, then prepare a patch for review. Codex is presented as part of that process, while humans and security researchers remain involved in deciding what to investigate and reviewing proposed fixes.

The 25% figure also illustrates the organizational commitment OpenAI says the prior sprint required. It should not be read as a recommended staffing ratio: the post gives no cost, output, or comparison data. Still, teams can take the underlying operational point seriously. Defensive use of coding models needs bounded environments and a review path, not just access to a capable model. Partnering with outside researchers may widen the range of approaches, although the announcement does not establish what Cathedral will find or whether the next sprint will outperform the last. Later coverage: [OpenAI: Research & Deployment](https://newruntime.com/posts/openai-research-deployment-7a8150/).

## Limits and uncertainty
This is a company announcement about planned work and OpenAI’s own account of a previous sprint. It does not establish when the new sprint will take place, what systems or models are in scope, or what Cathedral Systems will be able to access. OpenAI gives no results from the planned collaboration and no metrics for the previous sprint beyond the share of product engineers dedicated to it. The post also does not disclose bug bounty reward amounts, the number or severity of findings, fixes shipped, or independent validation of its claims. The $1 billion commitment concerns subsidized access for frontline defenders over six months; the announcement does not specify how much has been used or link it directly to the sprint.

## Skill to build now
A team can adapt the workflow OpenAI describes this week without treating model output as a verified vulnerability: route a candidate finding into an isolated reproduction environment, document the evidence needed to confirm it, and require review before a patch is merged. Keep the model’s permissions and test scope bounded. This is a practical process suggestion based on the announced workflow, not a claim that OpenAI has published a complete playbook.

## Reader takeaway
OpenAI is planning a security sprint with Cathedral Systems and says its previous sprint used Codex to help engineers investigate and reproduce vulnerabilities before patch review. The announcement offers a concrete workflow and a substantial staffing signal, but no outcome data for either sprint. Treat it as evidence of how one company is organizing defensive work, not proof that AI-assisted security has already delivered a measured advantage.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Codex](https://newruntime.com/topics/codex.md): Explore the codex topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)