The AI Brief

On June 22, OpenAI expanded its Daybreak cybersecurity program with a new open-source patching initiative called Patch the Planet, an updated Codex Security plugin, a partner program, and the full release of its most capable defensive AI model, GPT-5.5-Cyber. On CyberGym, GPT-5.5-Cyber reached 85.6% compared with 81.8% for the standard GPT-5.5; it also scored 39.5% on ExploitGym versus 25.95% for GPT-5.5, and 69.8% on SEC-bench Pro versus 63.1%.

Trail of Bits has dedicated security engineers to work full-time with Codex and GPT-5.5-Cyber across 19 open-source projects, and has already identified hundreds of security issues and merged dozens of patches, with many more still undergoing coordinated disclosure. More than 30 projects have committed to take part, with early participants including cURL, the Go project, Python, Sigstore, and pyca/cryptography. The disclosed findings are striking: GPT-5.5-Cyber identified security-relevant components across more than 30 million lines of the Linux kernel code, flagging potential security issues and then validating them dynamically, generating 8 kernel pointer information leak proof-of-concepts and 24 local privilege escalation exploits. Hundreds of issues were identified; those are the subset for which PoCs were automatically generated.

OpenAI's models also turned up a 23-year-old use-after-free flaw in OpenBSD's kernel. GPT-5.5 identified CVE-2026-8390, a WebAssembly vulnerability that Mozilla patched two days before Pwn2Own Berlin, prompting five of six registered Firefox entries to withdraw. No Firefox exploit was successfully demonstrated at the competition. These moves come as frontier models from Anthropic and OpenAI are accelerating vulnerability discovery, leaving software maintainers overwhelmed with an ever-increasing volume of bugs that need to be verified, triaged, and patched. While previously the challenge lay in finding vulnerabilities, the bottleneck has now shifted to patching them. The Daybreak Cyber Partner Program's initial partners include Cisco, Cloudflare, CrowdStrike, Palo Alto Networks, Check Point, Fortinet, IBM, Okta, SentinelOne, and Wiz. The full version of GPT-5.5-Cyber remains restricted to verified defenders through OpenAI's Trusted Access for Cyber program; general-purpose API access is not available.

Reflection AI will pay $150 million a month beginning July 1, 2026 through 2029 for immediate access to Nvidia's latest GB300 AI chips and supporting hardware across SpaceX's Colossus 2 data center near Memphis, Tennessee. The deal is worth up to $6.3 billion and either company has the option to end the contract with 90 days' notice after the first three months. The deal is smaller than SpaceX's deals with Anthropic and Google, which cost the companies $1.25 billion per month and $920 million per month, respectively.

The startup, which was founded in 2024 by two former Google DeepMind researchers, said the compute deal is one of the largest announced open AI infrastructure commitments to date. Nvidia invested $800 million in Reflection, which is now getting access to Nvidia chips purchased by SpaceX. Some investors have called Reflection the "DeepSeek of the West," per the Wall Street Journal.

Reflection used the compute deal, its first, to tout the value of its open-weight AI strategy, which it has pitched as an open-source alternative to closed frontier labs like Anthropic and OpenAI. Open-weight AI models have received more attention following the US government's ban of Anthropic's closed models, Fable and Mythos. Reflection said the agreement gives it additional computing capacity to accelerate what it calls "American open intelligence." The startup has not yet released a public frontier open-source model, but it has been building momentum with government and national security customers. The company is working with the Department of Energy's Genesis Mission and has been part of broader Pentagon AI efforts.

The draft guidelines clarify that complex systems made up of several AI components, including so-called agentic AI systems, must be assessed holistically. The guidelines state that, where multiple AI components interact and their combined outputs materially influence a decision in a high-risk use case, the system is to be treated as a single AI system. As a result, individual components cannot rely on the Article 6(3) filter in isolation unless they are genuinely separable and do not contribute to a high-risk purpose. Accordingly, even if a component performs only a narrow procedural or preparatory task, it may still be classified as high-risk where, as part of a complex or agentic AI system, it contributes to outputs that materially influence an Annex III use case.

The European Commission published the draft guidelines on May 19, 2026, and launched a public consultation. Following requests from stakeholders, the consultation deadline was extended to 23 July 2026. Under the updated compliance timetable, requirements for stand-alone high-risk AI systems are now due to apply from December 2, 2027, while obligations for high-risk AI systems embedded in products will apply from August 2, 2028. The guidelines are non-binding in the sense that final authoritative interpretation rests with the Court of Justice of the EU, but these are the most detailed materials the Commission has published on high-risk classification to date, and they signal how regulators are likely to approach enforcement.

The practical exposure is largest for enterprises running orchestrated agent pipelines in financial services (credit scoring, fraud triage), HR (CV screening, performance review), and healthcare routing. A system designed as four "narrow" sub-agents each below the classification threshold may now constitute a single high-risk system under the holistic rule. The final guidelines, expected end of 2026, are where this language will be locked in or revised.

First covered in Vol. I, No. 28 (June 22, 2026). When Anthropic launched Claude Fable 5 on June 9, it announced that Pro, Max, Team, and seat-based Enterprise subscribers would get access at no extra cost through June 22. Starting June 23, Fable 5 usage requires consumption credits at the model's standard API rates of $10 input / $50 output per million tokens. The complication: Fable 5 was offline from June 12 to approximately June 18 due to the US government export control directive, meaning subscribers effectively received 4–5 days of free access out of the advertised 13 days. Anthropic has not announced any extension of the complimentary window.

Anthropic has reportedly restored access to Claude Fable 5, but the model returned inside a compliance architecture that no enterprise has yet had time to properly assess. Four parties now hold different pieces of the access control picture: Anthropic, the US Department of Commerce, Congress, and the major cloud providers, and their positions aren't aligned. Restoration under an export control directive doesn't cancel the underlying compliance obligation. It restructures it.

Details of the Commerce Department's involvement and the strictness of the export control directive made a rapid full resolution look highly improbable. Prediction markets price 57% odds that Fable 5 returns to full unrestricted operation before July 1, 67% before July 10, and 75% before July 17. The restricted jurisdiction list — the document that would allow enterprises to determine their own exposure — has not been published by Commerce.

Disclosure: Anthropic, mentioned in this item, is the company that develops Claude, which generates this brief.