---
type: "post"
id: "openai-s-hacking-ai-found-dozens-of-holes-2cedd9"
slug: "openai-s-hacking-ai-found-dozens-of-holes-2cedd9"
title: "OpenAI’s Hacking AI Found Dozens Of Holes In Apple iPhones, Macs"
description: "OpenAI’s GPT-6 Cyber helped researchers find roughly 35 vulnerabilities across Apple products, including flaws that could enable device takeover through a malicious website."
retrieval_nugget: "OpenAI’s GPT-6 Cyber helped researchers find roughly 35 vulnerabilities across Apple products, including flaws that could enable device takeover through a malicious website. The report matters because it shows offensive security work becoming faster with frontier models, while also underscoring that humans still supply methods and verify findings."
published_at: "2026-11-18"
updated_at: "2026-11-18"
record_date: "2026-11-18"
date_kind: "published_at"
topics: []
entities: ["OpenAI","GPT-6 Cyber","Calif","Apple","GPT-6.1 Sol","XBOW","Daybreak"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "GPT-6 Cyber and Calif researchers found about 35 vulnerabilities across Apple devices, including flaws that could enable takeover through a malicious website."
prior_refs: ["https://newruntime.com/posts/openai-launches-gpt-6-cyber-for-approved-security-ec25c6/"]
source_url: "https://www.forbes.com/sites/thomasbrewster/2026/11/17/openai-gpt-cyber-finds-dozens-of-vulnerabilities-in-apple-tech"
source_urls: ["https://www.forbes.com/sites/thomasbrewster/2026/11/17/openai-gpt-cyber-finds-dozens-of-vulnerabilities-in-apple-tech"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:openai-s-hacking-ai-found-dozens-of-holes-2cedd9"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/openai-s-hacking-ai-found-dozens-of-holes-2cedd9/","markdown":"https://newruntime.com/posts/openai-s-hacking-ai-found-dozens-of-holes-2cedd9.md","json":"https://newruntime.com/posts/openai-s-hacking-ai-found-dozens-of-holes-2cedd9.json"}
markdown_url: "https://newruntime.com/posts/openai-s-hacking-ai-found-dozens-of-holes-2cedd9.md"
---

# OpenAI’s Hacking AI Found Dozens Of Holes In Apple iPhones, Macs

## Retrieval answer

OpenAI’s GPT-6 Cyber helped researchers find roughly 35 vulnerabilities across Apple products, including flaws that could enable device takeover through a malicious website. The report matters because it shows offensive security work becoming faster with frontier models, while also underscoring that humans still supply methods and verify findings.

## Summary

OpenAI’s GPT-6 Cyber helped researchers find roughly 35 vulnerabilities across Apple products, including flaws that could enable device takeover through a malicious website. The report matters because it shows offensive security work becoming faster with frontier models, while also underscoring that humans still supply methods and verify findings.

## What happened

Forbes reported on November 17, 2026, that OpenAI and cybersecurity company Calif had used GPT-6 Cyber over recent months to uncover about 35 weaknesses of varying severity across Apple hardware and software. The products named include iPhones, Macs, iPads, Apple Watches, Apple TVs and Vision Pro headsets. Some of the reported flaws could allow an attacker to take over an iPhone or Mac through a malicious website.

The researchers also used GPT-6 Cyber to code an exploit that could remove protections provided by Apple’s Lockdown Mode. Apple describes that mode as deliberately limiting iPhone functionality to reduce avenues of attack, according to the article. Calif CEO Thai Duong told Forbes that the latest OpenAI agentic hacking product is a step up from earlier models in how quickly it can help code exploits. He also warned that the capability can benefit defenders seeking to patch systems, but would be a liability in malicious hands.

Forbes says Calif told it that Apple fixed many vulnerabilities after being alerted. Apple had not responded to the publication’s request for comment at the time of the article. OpenAI confirmed that GPT-6 Cyber had found vulnerabilities for which Apple credited it, but did not comment further. The article does not give a vulnerability-by-vulnerability list, severity breakdown, affected software versions or patch dates.

The reporting places the Apple work alongside other recent security testing. XBOW tasked GPT-6 Cyber with finding issues in 65 open-source web apps with known vulnerabilities. It found that the newer model worked twice as fast as GPT-6.1 Sol at finding and exploiting a weakness. That speed came with a substantial cost difference: Forbes reports that one dollar spent on Sol would likely uncover the same number and kinds of exploits as $17 spent on GPT-6. At OpenAI’s November 5 launch, the company said GPT-6 Cyber had helped identify more than 2,000 bugs in Linux and developed successful hacks for 948 of them.

The article’s interviewees stress that this is an efficiency gain, not the removal of human expertise. Researchers still had to provide methodology and verify that findings were genuine. Calif’s Duong recommended that people facing targeted attacks continue to use Lockdown Mode, saying the discovery of a way around some protections does not negate the mode’s effectiveness.

## What the archive adds

New Runtime’s earlier coverage, [OpenAI launches GPT-6 Cyber for approved security teams](https://newruntime.com/posts/openai-launches-gpt-6-cyber-for-approved-security-ec25c6/), described the model’s initial rollout to select approved Daybreak Red customers through API and Codex. This is a new fact on that story: researchers have now reported a broad set of Apple-product findings, alongside evidence about exploit-development speed and cost.

## Trend signal

This does not yet fit a pattern New Runtime tracks. The report describes a model helping find and exploit vulnerabilities, but it does not establish a durable agent security boundary or a reusable runtime design. The findings are significant evidence about capability and risk; they are not, by themselves, evidence that a particular security architecture is becoming the industry’s standard.

## Why now

The report arrives after OpenAI’s November 5 launch of GPT-6 Cyber, when the company said the model had helped identify thousands of Linux bugs. The Apple findings extend the discussion from launch claims to work reported by researchers across a large consumer-device ecosystem. The timing is consequential because defenders and vendors must weigh a model’s accelerated vulnerability research against the possibility that similar speed could assist attackers.

## Connections to other events

New Runtime’s earlier coverage of [GPT-6 Cyber’s launch](https://newruntime.com/posts/openai-launches-gpt-6-cyber-for-approved-security-ec25c6/) established that OpenAI was making the model available to approved security teams. This report adds a concrete Apple-focused use case and a reported comparison of speed and cost against GPT-6.1 Sol. Together, the two events raise a practical question for security teams: how to make faster discovery translate into timely, verified fixes without treating model output as proof of a vulnerability.

## Why it matters

For teams building with AI-assisted security tools, the main implication is not that a model can replace penetration testers. The reporting points to a workflow in which the model can accelerate exploit development, while a human researcher still determines the method and checks whether the result is real. That distinction should shape how teams evaluate these systems: speed alone does not establish useful coverage, and an unverified finding is not a patch plan.

The cost comparison also argues for task-specific evaluation. Forbes reports that GPT-6 Cyber was twice as fast as GPT-6.1 Sol on the cited test, but that equivalent exploit discovery could cost substantially more. A team deciding whether to use a more capable model needs to compare verified results and total evaluation cost, not just elapsed time. For device owners, Calif’s stated advice is to keep Lockdown Mode enabled when facing targeted threats; the reported bypass of some protections does not mean the mode has no value.

## Limits and uncertainty

The roughly 35 vulnerabilities are reported as a combined total across hardware and software, but the article does not provide an itemized list, severity distribution, affected versions or dates when fixes shipped. Apple had not commented to Forbes, and OpenAI offered no further detail beyond confirming vulnerabilities for which Apple credited the model. The cost and speed comparison concerns a test involving 65 open-source applications with known vulnerabilities; it should not be read as a general measure of performance on all security work. The report also does not establish how many findings were independently reproduced or whether any were exploited in the wild.

## Reader takeaway

GPT-6 Cyber’s reported Apple findings show a meaningful acceleration in security research, not autonomous, self-validating hacking. Treat model-generated vulnerabilities as leads that need expert validation, and judge model choice by verified discoveries as well as speed and cost. For people at risk of targeted attacks, the report does not overturn Calif’s recommendation to use Lockdown Mode.

## Open question

How many of the reported Apple vulnerabilities have been independently reproduced, and what versions and patch dates are affected?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
