# OpenAI GPT-6 Cyber: AI model finds 81 Firefox flaws and 948 Linux kernel vulnerabilities with working exploits

> Help Net Security reports OpenAI's GPT-6 Cyber, a version of GPT-6 Astra trained to find software vulnerabilities, test how attackers could exploit them and help develop fixes. It reportedly found 81 previously unknown Firefox vulnerabilities rated high severity by Mozilla, 948 Linux kernel vulnerabilities with working exploits, and two zero-days in engineering software. Access is limited to approved Daybreak Red customers.

- **Published:** 2026-11-06T10:43:00Z
- **Last reviewed:** 2026-11-06T10:43:00Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-5406
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-5406.png
- **ID:** TL-2026-5406
- **Severity:** MEDIUM
- **Category:** THREAT_INTEL
- **Status:** MONITORING
- **Detections:** 9 (1 shown free below) · **IOCs:** 11

## Description

On 2026-11-06 Help Net Security reported the release of OpenAI's GPT-6 Cyber, a specialized version of GPT-6 Astra trained for vulnerability discovery, exploit validation and fix development. According to the article, the model found 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity, and 948 Linux kernel vulnerabilities for which working exploits were produced. In testing with Dragos on a mock power plant it reportedly identified two zero-days in engineering software, and a critical Brotli vulnerability that allows remote code execution was found while testing OpenAI's safety measures. On the ExploitGym benchmark GPT-6 Cyber scored 48.5% versus 32.1% for the predecessor GPT-5.6 Cyber.

No CVE identifiers, threat actors, IOCs or in-the-wild exploitation are stated for the GPT-6 Cyber findings; the article is the single source for these figures, and vendor-side details (affected Linux kernel versions, Firefox release, engineering software vendors) were not published in the material reviewed. The significance is the capability: automated, large-scale vulnerability discovery combined with working exploit generation compresses the window between bug discovery and weaponization, and the same capability class would be dual-use if it were obtained by an attacker.

OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework. This is a vendor capability classification, not a CVSS or exploitation severity. Access is initially through the API and Codex for selected, approved Daybreak Red customers, with planned extension to all program participants. Reported safeguards include monitoring for unauthorized actions, targeted threat detection, private safety processing, stolen-credential protections, blocking of high-risk cyber assistance, sandboxing and human oversight. On the ExploitGym Honeypot test the model reportedly avoided out-of-scope exploitation 99.93% of the time without safety instructions and 100% with them.

Context from earlier coverage: Daybreak Red launched around 2026-08-10 with GPT-5.6-Cyber, the first OpenAI model assessed at the High cyber threshold, which OpenAI said found two chainable Chrome V8 flaws (memory corruption plus heap sandbox escape, reported as CVE-2026-15903) that Google patched. This item is tracked for monitoring, not as an actively exploited vulnerability.

## MITRE ATT&CK

- T1595.002 Vulnerability Scanning
- T1587.004 Exploits
- T1588.006 Vulnerabilities
- T1588.007 Artificial Intelligence
- T1203 Exploitation for Client Execution

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**Mozilla Firefox Process Spawning Shell or Script Interpreter (AI-Discovered Firefox Flaw Exploitation)** (`TL-2026-5406-SPL-001`)

Detects Firefox (or its plugin-container helpers) spawning a command shell, script host or download utility on Windows or Linux. A successful browser memory-corruption exploit (T1203) normally ends with the compromised renderer launching a payload stage. TL-2026-5406 reports 81 previously unknown high-severity Firefox flaws found by OpenAI GPT-6 Cyber and a shrinking discovery-to-exploit window; no in-the-wild exploitation is reported, so this is behavioural monitoring for post-exploit child-process behaviour, not a signature.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1203, T1059.001, T1059.004
- **Data sources:** sysmon, sysmon_for_linux, edr_process_telemetry
- **Severity:** high
- **Confidence:** medium

### Splunk SPL

```spl
index=sysmon (sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" OR sourcetype="sysmon:linux") EventCode=1
| eval parent=lower(ParentImage), child=lower(Image)
| where match(parent,"(\\|/)(firefox|firefox-bin|plugin-container)(\.exe)?$")
| where match(child,"(\\|/)(cmd|powershell|pwsh|wscript|cscript|mshta|rundll32|regsvr32|bash|sh|dash|zsh|python[0-9.]*|perl|curl|wget|nc|ncat)(\.exe)?$")
| stats count min(_time) as firstTime max(_time) as lastTime values(CommandLine) as child_commands by host User ParentImage Image
| convert ctime(firstTime) ctime(lastTime)
```

### Microsoft KQL

```kql
DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessFileName in~ ("firefox.exe", "firefox", "firefox-bin", "plugin-container.exe", "plugin-container")
| where FileName in~ ("cmd.exe", "powershell.exe", "pwsh.exe", "wscript.exe", "cscript.exe", "mshta.exe", "rundll32.exe", "regsvr32.exe", "bash", "sh", "dash", "zsh", "python", "python3", "perl", "curl", "wget", "nc", "ncat")
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLine
```

### Sigma

```yaml
title: Firefox Spawning Shell or Script Interpreter
id: 3f6a1c52-8b4e-4d27-9a31-5c7e2d90b1a4
status: experimental
description: Detects Firefox or its helper processes launching a shell, script host or download utility, consistent with post-exploitation of a browser memory-corruption flaw (TL-2026-5406 AI-discovered Firefox vulnerabilities).
references:
    - https://www.helpnetsecurity.com/2026/11/06/openai-gpt-6-cyber-released/
author: AII-Detector — ThreadLinqs Intelligence
date: 2026-11-06
tags:
    - attack.execution
    - attack.t1203
    - attack.t1059.001
    - attack.t1059.004
logsource:
    category: process_creation
detection:
    selection_parent:
        ParentImage|endswith:
            - '/firefox'
            - '/firefox-bin'
            - '/plugin-container'
            - '\firefox.exe'
            - '\plugin-container.exe'
    selection_child:
        Image|endswith:
            - '\cmd.exe'
            - '\powershell.exe'
            - '\pwsh.exe'
            - '\wscript.exe'
            - '\cscript.exe'
            - '\mshta.exe'
            - '\rundll32.exe'
            - '\regsvr32.exe'
            - '/bash'
            - '/sh'
            - '/dash'
            - '/zsh'
            - '/python3'
            - '/perl'
            - '/curl'
            - '/wget'
            - '/nc'
            - '/ncat'
    condition: selection_parent and selection_child
falsepositives:
    - Browser helper applications or external-protocol handlers configured by the organisation
    - Developer or IT testing of Firefox native-messaging hosts
level: high

```

**Known false positives:**

- Enterprise Firefox policies launching helper apps (e.g. external protocol handlers) that start cmd.exe or sh
- Native-messaging hosts for password managers or SSO tools implemented as scripts

**Tuning notes:** Allowlist known native-messaging host command lines and approved protocol handlers via ParentCommandLine/CommandLine. Remove curl/wget/python from the child list on developer workstations if noisy. Raise priority when the child launches from a user profile or temp directory.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- Node.js, Bun or Envoy Runtime Spawning Shell or Downloader ([indicator] RCE Exploitation) (locked: unlock with TL Creds)
- Exploit Source Compiled in World-Writable Directory on Linux Host (Kernel Exploit Staging) (locked: unlock with TL Creds)
- OpenAI Codex CLI Agent Executed on Endpoint or Server (locked: unlock with TL Creds)
- Vulnerability Scanner User-Agents Against Web Servers (Automated Active Scanning) (locked: unlock with TL Creds)
- Exploit Code Download or Exploit Framework Execution on Linux Host (locked: unlock with TL Creds)
- Setuid Bit Set on File in World-Writable Directory (Post-Exploitation Privilege Persistence) (locked: unlock with TL Creds)
- Linux Unprivileged Parent Spawning Root Process Outside Known Privilege Brokers (Kernel LPE) (locked: unlock with TL Creds)
- Web Traffic to OpenAI API and ChatGPT Endpoints From Servers or OT Segments (AI Capability Use) (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sources

- [OpenAI's GPT-6 Cyber found 81 Firefox flaws earlier AI testing missed](https://www.helpnetsecurity.com/2026/11/06/openai-gpt-6-cyber-released/)
- [GPT-5.6-Cyber refuses security researchers' requests far less often](https://helpnetsecurity.com/2026/08/11/openai-gpt-5-6-cyber-model/)
- [OpenAI Launches Daybreak Red With GPT-5.6-Cyber, First Model to Hit 'High' Cyber Threshold](https://aiweekly.co/node/9665)
- [OpenAI: Expanding Daybreak as the cyber defense window narrows](https://www.openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/)
- [Can OpenAI's New Daybreak Tools Stop Cyber Threats?](https://technologymagazine.com/news/can-openais-new-daybreak-tools-stop-cyber-threats)
- [Firefox zero day: Mozilla says Claude Mythos found 271 bugs](https://dev.to/simon_paxton/firefox-zero-day-mozilla-says-claude-mythos-found-271-bugs-5e3g)
- [Firefox Maker Mozilla Says AI Could End the Era of Zero-Day Vulnerabilities](https://hothardware.com/news/mozilla-says-ai-could-end-zero-day-vulnerabilities)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-5406