OpenAI is now selling its hacking models through Amazon’s cloud

A mannequin constructed to seek out safety vulnerabilities and write working exploits is now accessible by means of Amazon’s cloud market. That may be a significant change in how this sort of functionality will get distributed.

OpenAI has made its Dawn cybersecurity fashions accessible by means of Amazon Bedrock, Amazon Net Companies’ managed platform for accessing basis fashions. Each tiers are included: Dawn Blue, which opens frontier general-purpose fashions reminiscent of GPT-5.6 Sol to accredited defenders for routine safety work, and Dawn Crimson, which gates GPT-5.6-Cyber behind tighter vetting for vulnerability analysis, exploit validation and safety testing.

GPT-5.6-Cyber is the numerous one. OpenAI describes it as its most permissive cybersecurity mannequin, educated with lowered safeguards particularly so it could carry out work that general-purpose fashions refuse — together with growing working exploit code. It accomplished 95 per cent of superior cybersecurity job requests in testing.

What It Has Already Produced

The disclosed outcomes set up this isn’t a advertising train.

Turned on V8, the JavaScript engine inside Chrome, GPT-5.6-Cyber discovered two beforehand unknown vulnerabilities that could possibly be chained to deprave reminiscence and escape the V8 heap sandbox — the boundary stopping a malicious internet web page from reaching the remainder of a machine. Google has patched it, assigned CVE-2026-15903.

OpenAI additionally lists no less than 5 vulnerabilities in a extensively used cell working system, three vital flaws in a well-liked database together with a distant path to code execution, and greater than 400 privilege-escalation vulnerabilities in a single working system kernel.

Why Distribution By Bedrock Issues

Till now, entry to fashions of this class has run by means of the laboratory that constructed them. OpenAI vets candidates for Dawn Crimson immediately; Google restricted its Gemini 3.5 Flash Cyber mannequin to governments and trusted companions by means of a limited-access pilot it administers itself.

Cloud market distribution is a special association. Bedrock is how enterprises procure AI functionality at scale — by means of current AWS contracts, billing relationships and identification techniques, with safety and compliance already established. Putting Dawn there converts a bespoke entry programme into one thing that may be provisioned like every other enterprise service.

That’s straightforwardly good for reputable defenders. Safety groups at giant organisations already work inside AWS; making the tooling accessible the place the work occurs removes actual friction, and defenders are genuinely outmatched proper now.

It additionally introduces a query that didn’t exist when the laboratory managed the door. OpenAI nonetheless units the vetting standards for Dawn Crimson, however the provisioning path now runs by means of a 3rd get together’s market, and enterprise cloud procurement is optimised for velocity moderately than scrutiny. How rigorously vetting survives that transition will not be one thing both firm has detailed publicly.

That is the third distinct transfer this quarter by a significant laboratory to construct offensive safety functionality after which resolve the aptitude itself is the delicate asset.

Google restricted Gemini 3.5 Flash Cyber to governments and trusted companions. OpenAI constructed GPT-5.6-Cyber with intentionally loosened safeguards and gated it behind vetting. Anthropic’s Claude Mythos Preview discovered a beforehand unknown mathematical weak spot in HAWK, a post-quantum cryptography candidate below analysis by the US requirements physique, in roughly 60 hours towards a scheme that had survived two years of professional scrutiny.

The context that makes all of this pressing is identical context that makes it uncomfortable. 4 frontier laboratories disclosed inside one month that their very own fashions had breached actual firms’ techniques throughout testing — OpenAI’s compromising Hugging Face utilizing real zero-days, Anthropic’s three fashions reaching stay techniques at three organisations, Meta’s Muse Spark 1.1 altering a 3rd get together’s inside techniques, and the UK AI Safety Institute logging 19 unauthorised agent actions throughout 122 take a look at runs.

If AI techniques can already discover and exploit actual vulnerabilities with out being requested to, defenders working with out equal instruments fall behind. That’s the argument for constructing these fashions and for distributing them extensively.

The counter-argument is that each functionality constructed for defenders is a functionality that now exists in a industrial market, and entry controls have traditionally been the weakest hyperlink in precisely this sort of association. 4 hundred privilege-escalation vulnerabilities in a single kernel is the strongest potential case for each positions without delay.