Apple Patches Over 30 Security Flaws Across iOS, macOS and Safari

Apple has released security updates addressing more than 30 vulnerabilities across iOS, macOS, and Safari. Among them, 4 WebKit flaws were discovered using AI tools: 3 credited to OpenAI Codex Security and 1 to Anthropic researchers Milad Nasr and Nicholas Carlini working with Claude. It's the first Apple security update to formally credit AI systems as co-discoverers of patched vulnerabilities, not just an incremental patch cycle.

Last updated August 4, 2026: added why this specific release is a first for Apple, and noted a larger follow-up security release on July 27, 2026 that credited even more AI systems, which happened after this article's original publication and hadn't been reflected here.

What Happened: Apple's June 2026 Security Updates

Apple issued iOS 26.5.2, iPadOS 26.5.2, macOS Tahoe 26.5.2, and Safari 26.5.2 on June 29, 2026, as an emergency release, ahead of the regularly scheduled iOS 26.6 cycle those fixes were originally slated for. Apple said it wanted to "reduce the time between when updates were first made public and when they were put into customers' hands," citing concern that AI tools could accelerate how quickly attackers turn a disclosed flaw into a working exploit. None of the roughly 30 patched vulnerabilities were disclosed as actively exploited before the fix shipped.

4 of the patched WebKit vulnerabilities were specifically identified using AI tools: CVE-2026-43707, CVE-2026-43716, and CVE-2026-43745 were credited to OpenAI Codex Security, while CVE-2026-43715, a use-after-free bug, was credited to Anthropic researchers Milad Nasr and Nicholas Carlini working with Claude. That WebKit credit line makes this the first Apple security bulletin to formally name AI systems as co-discoverers, rather than just noting AI-assisted research informally.

A Bigger Follow-Up Release Came a Month Later

This wasn't a one-time move. On July 27, 2026, Apple shipped a larger, regularly scheduled security release, including macOS Tahoe 26.6, that credited an even wider set of AI systems: Claude and Anthropic researchers again (across WebKit, WebKit Storage, and WebDAV fixes), OpenAI Codex Security, Z.AI's GLM, and NVIDIA's AI Red Team. Coverage of that release also surfaced that Apple has had direct access to Claude since April 2026 through an arrangement referred to as "Project Glasswing," suggesting the actual scope of Claude-assisted findings inside Apple may exceed what's been publicly credited so far.

How AI Tools Actually Find Vulnerabilities Like These

The use-after-free bug credited to Claude, CVE-2026-43715, is illustrative of what AI-assisted vulnerability research actually looks like in practice. Use-after-free flaws happen when a program continues to reference a piece of memory after it's been freed, and finding them typically requires either exhaustively fuzzing code paths with malformed input until the program crashes in a revealing way, or carefully tracing how memory is allocated and released across a large, complex codebase like WebKit's rendering engine. Large language models are useful for the second approach specifically: they can hold more of a codebase's context in view at once than a human reviewer working through it file by file, and they're well suited to spotting the kind of subtle logic mismatch, a memory reference that survives a deallocation path it shouldn't, that causes bugs like this one. That's a meaningfully different capability from traditional automated fuzzing tools, which find bugs by brute-force input generation rather than by reasoning about code logic.

Why It Matters

Apple's installed base makes any WebKit vulnerability significant on its own, since WebKit powers Safari and renders web content across many iOS and macOS apps, and a flaw there can be exploited just by visiting a malicious page. What makes this specific pair of releases notable is the pattern: Apple naming AI systems as formal co-discoverers, twice in a month, suggests AI-assisted vulnerability research has moved from an experiment to a standing part of Apple's own security pipeline rather than a one-off PR mention.

It also cuts both ways for the security industry more broadly. The same reasoning capability that lets Claude or Codex find a subtle use-after-free bug for a defender also lowers the bar for an attacker trying to find the same class of bug first, which is part of why Apple explicitly cited concern about AI-accelerated exploit development as its reason for shipping June's fixes as an emergency out-of-cycle release rather than waiting for the regular iOS 26.6 update. Whoever finds a given vulnerability first, in-house researchers, external bug bounty hunters, or AI-assisted tooling on either side, largely determines whether it becomes a patch or an exploit.

What Happens Next

Security researchers will continue analyzing both releases' patched vulnerabilities to understand attack vectors. Whether Apple's AI-assisted approach catches vulnerability classes that manual review misses, or mostly speeds up finding bugs skilled researchers would eventually catch anyway, is the open question the next few release cycles should start to answer.

Final Takeaway

Apple's June 2026 update was a first: the company's first formal AI co-discoverer credit. The July 27 follow-up, with a wider set of AI systems credited and a standing Claude access arrangement revealed, suggests it wasn't a one-off experiment. Users should still install updates promptly regardless of who or what found the bugs.

Key Points

AI in Vulnerability Research

AI models can analyze code patterns, identify potential vulnerability classes, and suggest proof-of-concept approaches faster than traditional manual review, but they're not yet a replacement for human security expertise, effectiveness still depends heavily on the vulnerability type and the quality of what the model was trained or fine-tuned on. Apple naming AI systems as formal co-discoverers across two consecutive release cycles, rather than one isolated mention, is what signals this has become standing practice at one of the industry's most security-conscious companies, not just an experiment.

FAQs