AgentsAutonomy & tool use
Block moves security operations onto OpenAI's Daybreak and cites an 18-minute path from vulnerability report to proposed patch
Block, the company behind Square and Cash App, runs vulnerability discovery, detection testing, threat investigation and response on OpenAI's gated Daybreak stack. It says red-team exercises produced proposed patches within minutes and that it recovers hundreds of analyst-hours each quarter.
From the OpenAI News article of October 14, 2026 and cited sources: Block says Daybreak tooling went from a high-level vulnerability report to a proposed patch in 18 minutes during internal red-team exercises. [1] [2]
SOURCES[1] How Block is building AI-powered security with OpenAI… openai.com[2] Block AI-Powered Security Moves From Alerts to Patches in… remio.ai
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Block says Daybreak tooling went from a high-level vulnerability report to a proposed patch in 18 minutes during internal red-team exercises. [1] [2]
- Original source
- OpenAI News article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 14, 2026 article published
- Covered
- Oct 15 edition
- Evidence
- 7 sources across 6 sites
- 1How Block is building AI-powered security with OpenAI…openai.com
- 2Block AI-Powered Security Moves From Alerts to Patches in…remio.ai
- 3Block CISO: We red-teamed our own AI agent (The Register)theregister.com
- 4OpenAI Launches Daybreak for AI-Powered Vulnerability…thehackernews.com
- 5OpenAI Expands Daybreak With GPT-5.5-Cyber (The Hacker News)thehackernews.com
- 6Block says OpenAI tools speed security response; shares…blockchain.news
Reach: not measured yet
Block says Daybreak tooling went from a high-level vulnerability report to a proposed patch in 18 minutes during internal red-team exercises. It also cites about 1,100 analyst-hours of investigation completed per quarter, about 600 hours per quarter freed from false-positive triage, and a 50% rise in triage throughput. Humans still decide which fixes ship. [1] [2]
Daybreak pairs GPT-5.5 tiers with Codex Security. Codex Security builds a threat model that analysts can edit, tests candidate vulnerabilities in an isolated sandbox and then proposes fixes. Block reportedly picks a model for each task based on job type, risk and its own evaluations, and can replay seven months of telemetry in under an hour to test detection changes. [2] [4]
Earlier, Block's offensive security team compromised its own goose agent. Prompt injections hidden in invisible Unicode, delivered through calendar invites and shared workflows, got the agent to run an infostealer on an employee's laptop. Block responded by stripping those characters and adding prompt-injection detection. [3] [7]
The evidence shows a large fintech running AI across its security workflow while humans approve fixes. Faster triage and patching would follow only if the figures hold, and no outside party has checked them.
Read the full assessment
Patching capacity, not discovery, may remain the bottleneck.
Executive brief
Block says that in recent red-team exercises, OpenAI-based tooling went from a high-level vulnerability report to an automatically proposed patch in 18 minutes (Remio). Block, the company behind Square and Cash App, runs its security workflow on Daybreak, OpenAI's governed cyber-defense stack. Block uses it to find vulnerabilities, write and test detections, investigate threats and respond. Block also reports getting back about 600 analyst-hours per quarter from false-positive triage (OpenAI). All of these figures come from the companies themselves. Nobody has audited them, and key denominators are missing.
What changed and event timeline
Block red-teams its own agent
Block's CISO described an internal exercise in which a prompt injection hidden in invisible Unicode got Block's goose agent to run an infostealer on an employee's laptop ().
OpenAI launches Daybreak
Daybreak pairs GPT-5.5 tiers with Codex Security for threat modeling, sandboxed vulnerability validation and proposed fixes. Access is tightly controlled ().
GPT-5.5-Cyber expands
OpenAI calls the updated GPT-5.5-Cyber its strongest model for finding and patching vulnerabilities. It is released only to trusted defenders, alongside the "Patch the Planet" open-source effort ().
OpenAI publishes the Block case study
The case study says Block's security team investigates threats faster and proposes patches in minutes using OpenAI technology ().
Muted market reaction
Block shares closed down 1.31% at $75.85. Coverage noted that the disclosure gave little detail on usage or scope ().
Capabilities and access
Organizations get access by requesting it through OpenAI (The Hacker News).
Read the full section
- Stack: Daybreak, which combines frontier models, the Codex harness, security tools, access controls and human oversight (Remio).
- Exact model: none of the reviewed sources names one for Block. Block reportedly picks a model for each task based on job type, risk and its own internal evaluations, rather than always using the newest one (Remio).
- Daybreak model tiers:
- GPT-5.5 with standard safeguards
- GPT-5.5 with Trusted Access for Cyber
- GPT-5.5-Cyber for red teaming
Organizations get access by requesting it through OpenAI (The Hacker News).
Technical analysis for researchers and developers
- Patch pipeline: Codex Security builds a threat model of a repository that analysts can edit.
- Detection engineering: Block reportedly replays seven months of production telemetry in under an hour to test detection changes.
- Evaluation and reproducibility: there is no public benchmark, dataset or methodology. The 18-minute figure comes from internal red-team exercises, not production incidents.
Read the full section
- Patch pipeline: Codex Security builds a threat model of a repository that analysts can edit. It then tests candidate vulnerabilities in an isolated environment before proposing fixes (The Hacker News).
- Detection engineering: Block reportedly replays seven months of production telemetry in under an hour to test detection changes. That used to take days or weeks (Remio).
- Evaluation and reproducibility: there is no public benchmark, dataset or methodology. The 18-minute figure comes from internal red-team exercises, not production incidents.
- Implementation takeaway: routing each task to a model based on risk and internal evals is the transferable design pattern.
Claims and evidence
All of these figures are reported by Block or OpenAI in a vendor case study. Independent corroboration is not available. Reporting also conflicts on detail. Remio describes the Daybreak stack and gives these metrics, while Blockchain.News says the disclosure identified neither the model nor the time savings.
Read the full section
| Claim | Source |
| 18 minutes from vulnerability submission to proposed patch | Remio |
| About 1,100 analyst-hours of investigation completed per quarter | OpenAI |
| About 600 analyst-hours per quarter returned from false-positive triage | OpenAI |
| 50% increase in triage throughput | Remio |
All of these figures are reported by Block or OpenAI in a vendor case study. Independent corroboration is not available. Reporting also conflicts on detail. Remio describes the Daybreak stack and gives these metrics, while Blockchain.News says the disclosure identified neither the model nor the time savings.
Context and prior work
- Block's own agent security: Block's offensive security team compromised its own goose agent using zero-width Unicode prompt injections delivered through calendar invites and poisoned "recipes" (shareable workflows).
- Competing approach: Anthropic's Mythos initiative has a similar aim of finding and fixing flaws before attackers do (The Hacker News).
- Daybreak's other partners: CrowdStrike, Palo Alto Networks, Cloudflare and others (The Hacker News).
Read the full section
- Block's own agent security: Block's offensive security team compromised its own goose agent using zero-width Unicode prompt injections delivered through calendar invites and poisoned "recipes" (shareable workflows). Fixes included stripping those characters and adding prompt-injection detection (Block Engineering, The Register).
- Competing approach: Anthropic's Mythos initiative has a similar aim of finding and fixing flaws before attackers do (The Hacker News).
- Daybreak's other partners: CrowdStrike, Palo Alto Networks, Cloudflare and others (The Hacker News).
Limitations, safety and contested findings
- Missing denominators: the public case study does not give the total number of alerts, how often analysts overturned AI conclusions, false-negative rates, or how many proposed patches reached production (Remio).
- Agent attack surface: Block's goose exercise shows agents with tool access can be turned against their own users (The Register).
- Patching bottleneck: coverage notes that AI is speeding up vulnerability discovery faster than organizations can patch (The Hacker News).
Read the full section
- Missing denominators: the public case study does not give the total number of alerts, how often analysts overturned AI conclusions, false-negative rates, or how many proposed patches reached production (Remio).
- Agent attack surface: Block's goose exercise shows agents with tool access can be turned against their own users (The Register).
- Patching bottleneck: coverage notes that AI is speeding up vulnerability discovery faster than organizations can patch (The Hacker News).
Business and practitioner implications
- Where the gains are: the reported gains are mostly in analyst time (triage, investigation, detection testing), not in replacing analysts.
- Access: Daybreak and GPT-5.5-Cyber are gated, so smaller teams may not get the same tooling quickly (The Hacker News).
- If you deploy something similar:
Read the full section
- Where the gains are: the reported gains are mostly in analyst time (triage, investigation, detection testing), not in replacing analysts. Patches are proposed, and humans still decide what ships.
- Access: Daybreak and GPT-5.5-Cyber are gated, so smaller teams may not get the same tooling quickly (The Hacker News).
- If you deploy something similar:
- Track reversal rates, false negatives and how many proposals actually merge before claiming ROI.
- Red-team your own agents against prompt injection, as Block did.
- Market signal: investors did not react much to the announcement (Blockchain.News).