OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
GPT-6 Cyber, OpenAI's new model for finding vulnerabilities, worked with security firm Calif to uncover about 35 flaws in iPhone, Mac and other Apple hardware. Some of them could theoretically let an attacker take over a device if the victim opens a malicious website. For the iPhone, the team built an exploit (ready-to-run attack code) that bypasses Lockdown mode, which deliberately limits the phone's features. Apple has already closed many of the flaws, so security teams should speed up patch deployment.
AI-processed from 3DNews AI; edited by Hamidun NewsImagine an employee who finds dozens of breaches in the most secure phones and computers in the world within a few weeks. For defenders, that is a gift. In the wrong hands, it is a weapon. That is the story OpenAI and cybersecurity firm Calif have told: their new vulnerability-hunting model, GPT-6 Cyber, found about 35 vulnerabilities of varying severity in Apple's hardware and software. Among them is a way around Lockdown, the iPhone's strict protection mode that deliberately restricts device features so attackers have fewer openings.
What They Found
Over recent months, OpenAI and its partner Calif ran Apple products through GPT-6 Cyber. Nearly every product line was targeted: iPhone, Mac, iPad, Apple Watch, Apple TV and Vision Pro headsets. The figure "about 35" is the number of vulnerabilities found, ranging from less dangerous to serious.
Not all findings are equally alarming. According to the source, some could theoretically give an attacker control over a Mac or iPhone through a malicious website. Put simply, the victim only needs to open a specially prepared page. That is an unpleasant scenario because it requires no app installation.
Separately, the researchers used the model to create an exploit capable of neutralizing Lockdown mode's protections on the iPhone. An exploit is ready-made code that turns a discovered vulnerability into a working attack. Lockdown mode exists precisely to reduce the number of possible attack vectors, so bypassing it is a notable result. The source does not disclose details.
Calif said Apple fixed many of the discovered vulnerabilities shortly after receiving the information. How many flaws remain open is unknown.
Why Speed Matters
Calif CEO Thai Duong (Thai Duong) told Forbes that GPT-6 Cyber significantly outperforms its predecessors at creating exploits. Previously, this stage, from discovering a flaw to building a working attack, required specialists a lot of time and skill.
According to Duong, this speed cuts both ways. It helps defenders find and close holes faster. But if attackers gain access to the model, it could become a serious threat.
Calif is not new to using AI to find bugs. The source mentions a memory-management bug in Java called Java4Shell. In August, the company found a vulnerability in PHP, but that time it used a different model, GPT-5.6 Sol. A clear trend emerges: with each model generation, more flaws are being found in long-studied, widely used code.
What It Means in Practice
First, the window between a patch's release and the start of mass attacks will likely shrink. If exploits can be written faster, attackers need less time too. Rapid rollout of updates becomes a requirement, not a nice-to-have.
Second, teams responsible for their own products should treat such models as an auditing tool. If AI finds dozens of flaws in products from a company of Apple's scale, ordinary corporate code is unlikely to have fewer. Running code through such tools before release is a sensible practice.
Third, access to these models becomes a security question. Who can use them, and under what conditions, determines whether they help defenders or attackers more.
Limitations
The source is a brief 3DNews note citing Forbes, so technical details are scarce. It is unknown which specific vulnerabilities were found, how easily they can be exploited, or how much of the work was done by the model versus Calif's specialists. The word "theoretically" in the description of web attacks also matters: it means the source contains no facts about real attacks on users. Still, the direction is clear: AI for cybersecurity is already strong enough to find flaws in some of the most protected consumer platforms.
- Nov 21, 2026California Sues OpenAI: Its AI Agents Escaped Control at Least 80 Times
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
- Nov 21, 2026Google to split Gemini chat history from model training, so you can opt out without losing your history
- Nov 21, 2026Apple's macOS 27.3 beta 2 will let you limit AI agents' access to Mac files
Key Questions Answered
What They Found?
Over recent months, OpenAI and its partner Calif ran Apple products through GPT-6 Cyber. Nearly every product line was targeted: iPhone, Mac, iPad, Apple Watch, Apple TV and Vision Pro headsets. The figure "about 35" is the number of vulnerabilities found, ranging from less dangerous to serious. Not all findings are equally alarming. According to the source, some could theoretically give an attacker control over a Mac or iPhone through a malicious website.
Why Speed Matters?
Calif CEO Thai Duong (Thai Duong) told Forbes that GPT-6 Cyber significantly outperforms its predecessors at creating exploits. Previously, this stage, from discovering a flaw to building a working attack, required specialists a lot of time and skill. According to Duong, this speed cuts both ways. It helps defenders find and close holes faster. But if attackers gain access to the model, it could become a serious threat. Calif is not new to using AI to find bugs.
What It Means in Practice?
First, the window between a patch's release and the start of mass attacks will likely shrink. If exploits can be written faster, attackers need less time too. Rapid rollout of updates becomes a requirement, not a nice-to-have. Second, teams responsible for their own products should treat such models as an auditing tool. If AI finds dozens of flaws in products from a company of Apple's scale, ordinary corporate code is unlikely to have fewer. Running code through such tools before release is a sensible practice.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.