OpenAI Next Cyber Model: 77% ExploitGym Teaser Claimed vs Verified

Coverage of OpenAI’s November 12, 2026 Cyber at the Frontier session includes a briefly shown chart that placed a “next cyber model” at 77.0% on ExploitGym — roughly 28 points above the 48.5% score OpenAI documents for GPT-6 Cyber. OpenAI’s own Cyber at the Frontier page markets Daybreak Defense Factory and Red Team, not a new public cyber base model. Until OpenAI posts a system card or product name, treat 77% as a livestream teaser, not a shipped capability.

TL;DR: claimed vs verified

What ExploitGym measures

ExploitGym tests whether a model can exploit a specified vulnerability in a remote target and recover a flag, proving the intended weakness was used. OpenAI uses it alongside honeypot evaluations modeled on the Hugging Face autonomous agent breach to measure out-of-scope behavior.

Our GPT-6 Cyber launch post explains why a few point swing matters: sub-50% success still implies routine exploitation on scoped tasks at API prices. A 77% line, if it ever ships, would move procurement conversations for Daybreak Red — but only after OpenAI documents hardened versus standard ExploitGym variants (the system card also lists a hardened track near 37% for GPT-6 Cyber).

What the livestream reportedly showed

Gizmodo Japan’s write-up says moderators displayed a benchmark graphic that already included GPT-6 Astra and GPT-6 Cyber, then added a red line for an unreleased cyber-specialized model at 77.0%. The article contrasts that with Cyber’s published 48.5% and cites an on-slide Anthropic Mythos figure near 38% — we treat Anthropic’s number as slide-only unless Anthropic republishes it.

The Cyber at the Frontier page itself is a CTA shell: Defense Factory, Red Team, and Request Daybreak Access. That matches explainx.ai’s read that November 12 was go-to-market for defender workflows, not necessarily a second cyber foundation model release.

OpenAI next cyber model ExploitGym teaser versus verified GPT-6 Cyber scores

What OpenAI already shipped this week

Official November 12 artifacts align with agentic security products:

• Defense Factory — continuous scan, triage, patch proposals.
• Red Team — authorized app testing agents.
• Developer blog on scaling defenders tying Codex Security and SARIF ingestion to the same program.

Those are the pieces security engineers can request access to today. They sit on top of GPT-6 Cyber and Daybreak Blue/Red models described on November 5, not the teased 77% line.

For comparison baselines, read GPT-6 Cyber vs GPT-5.6 Cyber and fuzzy robustness red-teaming on GPT-6 Astra.

Why vendors flash unreleased curves

Cyber models carry Preparedness Framework labels. OpenAI classified GPT-6 Cyber as Critical in cybersecurity. Previewing a higher ExploitGym point on a livestream signals research momentum to Daybreak Red prospects without triggering the same deployment review paperwork as a named GA model.

That is standard frontier-lab behavior — similar to how Linux kernel mass-exploit posts referenced an “upcoming cyber model” before the November 5 name drop. The difference for readers is documentation: November 5 brought PDFs; November 12’s 77% flash did not yet.

ExploitGym variants matter for any new score

OpenAI’s system card distinguishes standard and hardened ExploitGym tracks. GPT-6 Cyber lands near 48% on the standard track and about 37% on hardened targets with stronger mitigations.

Livestream observers rarely capture which variant a teaser line used. A 77% point could reflect:

• A less hardened target set
• Additional deployment prompts not shipped to Red customers yet
• An internal checkpoint absent alignment training

Without methodology footnotes, comparing 77% to the 48.5% headline is apples-to-oranges risk. The same caution applied when vendors swapped SemiAnalysis GDPx panels without publishing task counts.

Daybreak program timeline for context

The teaser fits OpenAI’s pattern of showing forward curves to Red prospects between numbered releases, not replacing the last shipped card overnight.

Preparedness Framework and access reality

Critical cybersecurity rating implies models can develop functional zero-days on hardened systems without humans in the loop, per OpenAI’s framework blog lineage traced in Astra Critical disclosure.

Access remains tightly gated:

• Default Daybreak Red networking restrictions
• Human approval on destructive chains
• Separate Blue tier for lower-risk validation work

A higher ExploitGym screenshot does not automatically move those gates. Regulatory readers should also watch parallel litigation threads such as the California AI freeze complaint — unrelated legally, but part of the same news cycle as Cyber at the Frontier.

How Red teams should interpret teaser charts

Authorized testers on Daybreak Red already juggle:

• Scope prompts that constrain hosts and CVE ids
• Evidence requirements before merge
• Cost caps on long-horizon exploit chains (Linux kernel run cited $1.2M API spend)

A higher ExploitGym point might shorten time-to-flag on scoped labs without changing authorization paperwork. Until OpenAI documents inference cost and refusal rates for the teased checkpoint, do not re-baseline SLAs.

Pair qualitative testing with fuzzy robustness red-teaming on mainline models you already run — most production risk still sits in misconfigured agents, not unreleased cyber weights.

Media coverage map (secondary sources)

No secondary source replaces a system card. Use them to triangulate whether a teaser became a product.

Procurement language to use today

When writing RFP text, cite documented figures:

• “Vendor shall disclose intended-exploit success on ExploitGym standard track with published methodology.”
• “Vendor shall provide out-of-scope rates on HoneyPot-style evaluations.”
• “Any demo score not yet on a system card is non-binding.”

That keeps security teams from budgeting headcount against a livestream line that may reflect internal research only.

What defenders should do now

1. Budget to shipped numbers — plan purple-team capacity using 48.5% class capability with known scope prompts, not speculative 77%.
2. Pilot Daybreak workflows — Factory and Red Team integrations matter more for most enterprises than incremental ExploitGym points.
3. Keep honeypot evaluations — OpenAI’s under 1% out-of-scope stat on HoneyPot is as operationally important as raw exploit success; see alignment limits research for why scope adherence decays under pressure.
4. Track access tier — GovCloud Daybreak approvals still gate who can run Critical-rated models.

Verified Daybreak cyber defenses versus unverified next-model ExploitGym teaser

What to watch for verification

OpenAI would confirm the teaser by publishing at least one of:

• A named model on openai.com/index/ with ExploitGym methodology footnotes
• An updated system card contrasting standard versus hardened tracks
• Daybreak Red release notes with a versioned cyber model ID
• Pricing or Preparedness tier change tied to the new score

Until then, cite 77% only as livestream-reported, link Gizmodo or primary video if OpenAI posts it, and anchor procurement docs to 48.5%.

Related reading

• GPT-6 Cyber launch and HoneyPot scores
• GPT-6 Cyber vs GPT-5.6 Cyber comparison
• Daybreak Defense Factory and Red Team
• Hugging Face breach and honeypot design
• OpenAI fuzzy robustness red teaming

Open questions for the next OpenAI security post

Builders should watch whether OpenAI (a) names the teased checkpoint, (b) separates Red-only weights from anything approaching mainline APIs, and (c) updates HoneyPot out-of-scope rates alongside ExploitGym lifts. A model that exploits faster but wanders off scope more often is a net loss for enterprise purple teams, regardless of headline percentage.

If the company instead ships only Defense Factory workflow polish, the 77% line may remain an internal research footnote — the same way pre-November kernel runs referenced an unnamed cyber stack before GPT-6 Cyber documentation caught up.

Subscribe to OpenAI’s deployment safety mailing list and archive Daybreak release notes if your team can legally hold them; that is usually where out-of-band score updates land before marketing blogs. Save livestream timestamps and screenshots if your compliance team needs an audit trail proving you did not plan around unverified capability. Forward those archives to your GRC queue today so procurement and security share one source of truth.

If OpenAI publishes the next cyber model formally, we will update this post with official scores and access paths.