共有:

OpenAI · Safety Governance

GPT-5.6 Sol's headline relaunch
stalls under three simultaneous alarms.

OpenAI only reopened GPT-5.6 Sol worldwide on July 9. Within days, the head of safety resigned, the UK's AI Security Institute confirmed a universal jailbreak, and OpenAI itself admitted ChatGPT Work isn't ready — three alarms in one window. Everyday chat users see little change on the surface, yet enterprise approval memos will now carry a new line: "safety review pending."

AI Navigate Editorial2026.07.127 min read

FIVE DAYS AFTER RELAUNCH 7/9 7/10 7/11 7/12 Sol relaunch Safety chief exits Head of Safety Universal JB found UK AISI Work admits gap OpenAI self Approval memos gain a "safety review pending" line
01

What Happened

Three alarms fired
inside a 72-hour window

Right after the worldwide GPT-5.6 Sol relaunch, three independent events collided.

First, news broke that OpenAI's Head of Safety had resigned. GPT-5.6 Sol had only been cleared by US regulators for global re-availability on July 9, and losing the most senior safety-governance executive that same week set off internal-friction rumors.

Second, the UK government-backed lab UK AI Security Institute (AISI) reported it had found a universal jailbreak against GPT-5.6 Sol. "Universal" here means the failure is not tied to a particular topic — the model's refusal policy can be sidestepped across the board.

Third, OpenAI itself acknowledged, through a senior executive, that ChatGPT Work — the new enterprise tier layered on top of Business and Enterprise — is not yet feature-complete. The three events are independent, but from a procurement seat they line up: flagship model, safety governance, and enterprise SKU all look shaky at once.

02

By The Numbers

Names and numbers
to keep in mind

7/9
GPT-5.6 Sol relaunch
72h
span of the three alarms
Universal
character of the AISI JB
Sol / Terra / Luna
three GPT-5.6 tiers
Work
SKU admitted as unfinished
Head of Safety
role that resigned
03

Why It Matters

Not three coincidences —
a governance seam

If they had fired at random, you could shrug it off. Firing at once shifts the meaning.

Frontier-lab safety has been read as a three-layer stack: model-level red-teaming, external evaluation, and board-level governance. This week each layer flashed a warning independently: no head of safety at the top, a universal vulnerability from external evaluators, and a self-admitted SKU that isn't ready. Three at once is not "an oops" — it reads as a mismatch between release velocity and safety-review capacity.

Since late June OpenAI has been countering Claude's enterprise inroads by shipping SKUs and models on parallel tracks. This 72-hour window is the first case where the cost of the "ship fast" strategy has become visible from the buyer's side.

04

Who It Hits

Who feels this, and how

RolePractical impact from 7/12 on
IT / ProcurementAdd "safety review pending" to approval memos. Line up Claude and Bedrock as concrete alternates.
Product / EngDelay ChatGPT Work-dependent PoCs by 1–2 weeks. Re-check rollback paths and MCP tool permissions.
SecurityWhen AISI publishes the JB method, run it against your own agent stack to measure exposure.
ExecutiveSingle-vendor lock-in is now not only a performance risk but a governance-linked risk.
ConsumerDay-to-day ChatGPT usage is unchanged. Reconfirm what sensitive data you paste.

A flagship's quality and its
governance can drift out of sync.


05

What To Do Next

What's next,
and what to do this week

Three recommended actions and the near-term outlook.

01

Audit MCP tool permissions

Whether a universal JB reaches production depends on which tools the agent can call, not on the model alone. This week, put human approval back in front of write-side tools (databases, mail, deploys) to buy margin.

02

Fall back to Bedrock for the same models

Even while the top-tier SKU sits under review, GPT-5.6 Sol / Terra / Luna via Amazon Bedrock stay inside AWS's audit boundary. Move audit-heavy or region-bound workloads off direct API to Bedrock to keep approvals moving.

03

Push ChatGPT Work back a month

A self-admitted "not ready" is not a weekly-patch problem. Requirements Business or Enterprise can cover today should stay there; re-evaluate Work in late August or early September.

06

Counterview

Counterview and limits

The reading here has weaknesses. The resignation, the AISI finding, and the Work admission may share a timeline without sharing causation — closeness in time is not causation. And "universal" is doing a lot of work in one word; the severity depends on the AISI report's scope and method, which will only be clear when the full document is published.

A wider frame also helps: the industry has seen three similar safety-lead departures in the past twelve months. That points to a frontier-wide tension between release cadence and safety capacity rather than an OpenAI-only failure. This week's three alarms are best read as that industry-wide pressure showing up at one lab in one window.