# OpenAI GPT-6 Cyber AI Model Finds 81 Firefox Flaws and Develops Root Exploits for 948 Linux Kernel Flaws

> OpenAI unveiled GPT-6 Cyber, a model for vulnerability discovery and exploit development rated 'Critical' under its Preparedness Framework. Per BleepingComputer it found 81 previously unknown high-severity Firefox vulnerabilities and produced local privilege escalation exploits for 948 distinct Linux kernel flaws. Access is limited to approved customers in OpenAI's Daybreak Red program.

- **Published:** 2026-11-06T20:44:01Z
- **Last reviewed:** 2026-11-06T20:44:01Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-5445
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-5445.png
- **ID:** TL-2026-5445
- **Severity:** MEDIUM
- **Category:** THREAT_INTEL
- **Status:** MONITORING
- **Detections:** 9 (1 shown free below) · **IOCs:** 14

## Description

This record tracks an emerging capability, not an exploited vulnerability: AI-accelerated vulnerability discovery and exploit development. No CVE, in-the-wild exploitation, threat actor, or network indicator is reported in any source.

BleepingComputer (Lawrence Abrams, 2026-11-06) reports that OpenAI's GPT-6 Cyber is rated 'Critical' under the OpenAI Preparedness Framework, up from the 'High' rating of the previous model. The article states the model found 81 previously unknown high-severity Firefox vulnerabilities and links the work to Mozilla advisory MFSA2026-118. In Linux kernel research it identified 2,308 potential issues, reproduced 1,724, and developed local privilege escalation exploits (unprivileged user to root) for 948 distinct flaws. On the ExploitGym benchmark it reached a 48.5% exploit success rate versus 32.1% for GPT-5.6 Cyber. Access is not public; it is limited to approved Daybreak Red customers, with expansion to all Daybreak Red customers planned. OpenAI is quoted: 'It is clear that we have a limited window to use frontier models to secure software.'

Corroboration: Mozilla's advisory MFSA2026-118 (Firefox 159, announced 2026-10-27) fixes well over 100 CVEs (repeated automated extractions of the page returned inconsistent totals, so no exact count is asserted here) and credits OpenAI Codex Security (handle 'kylebot', dozens of entries) and Amy Burnett of OpenAI, with most findings identified internally by Mozilla. That advisory does not itself confirm the figure of 81 or attribute the findings to GPT-6 Cyber, so the 81 and 948 figures rest on the single BleepingComputer article. Secondary sources describe the Daybreak program: Daybreak Blue gives approved defenders GPT-5.6 Sol without system-level cyber safeguards, and Daybreak Red offers purpose-trained cyber models (GPT-5.6-Cyber, announced 2026-08-10, rated 'High') for authorized vulnerability research and exploit validation, gated by identity verification, account monitoring and legal attestations. Some secondary sources attribute the first 'Critical' cyber rating to a separate model, GPT-6 Astra (2026-09-03/04); the relationship between GPT-6 Astra and GPT-6 Cyber is not established in the reviewed sources.

Defensive relevance: the time from disclosure to a working Linux LPE exploit is likely to shrink as comparable capabilities spread. Defenders should prioritize rapid kernel and Firefox patching, monitor for unprivileged-to-root escalation behavior, and track whether comparable capability reaches unrestricted or adversary-accessible models. Severity and status are analyst-assigned, not stated by the sources.

## MITRE ATT&CK

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

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

**Unprivileged-to-Root Shell via Kernel LPE (auditd euid 0 / uid non-zero)** (`TL-2026-5445-SPL-002`)

TL-2026-5445: GPT-6 Cyber reportedly produced unprivileged-to-root Linux kernel LPE exploits for 948 flaws. Detects a shell whose effective uid is 0 while its real uid is a logged-in non-root user, the credential-transition result of kernel exploitation and setuid abuse (T1068, T1548.001).

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1068, T1548.001
- **Data sources:** linux_auditd, index=linux sourcetype=linux:audit, DeviceProcessEvents
- **Severity:** critical
- **Confidence:** high

### Splunk SPL

```spl
index=linux sourcetype="linux:audit" type=SYSCALL success=yes euid=0 uid!=0 auid!=0 auid!=4294967295 (exe="/bin/sh" OR exe="/bin/bash" OR exe="/bin/dash" OR exe="/usr/bin/sh" OR exe="/usr/bin/bash" OR exe="/usr/bin/dash" OR exe="/usr/bin/zsh")
| stats count min(_time) as first_seen values(comm) as comm values(ppid) as ppid by host auid uid euid exe
| eval first_seen=strftime(first_seen,"%Y-%m-%d %H:%M:%S")
| table first_seen host auid uid euid exe comm ppid count
```

### Microsoft KQL

```kql
DeviceProcessEvents
| where Timestamp > ago(24h)
| where AccountName == "root" and InitiatingProcessAccountName != "root"
| where FileName in~ ("sh", "bash", "dash", "zsh")
| where InitiatingProcessFileName !in~ ("sudo", "su", "login", "sshd", "pkexec", "cron", "crond", "systemd", "gdm-session-worker", "lightdm")
| project Timestamp, DeviceName, InitiatingProcessAccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLine, ReportId
```

### Sigma

```yaml
title: Linux Shell Running With Effective UID 0 and Non-Root Real UID
id: 7d2e9a54-5b13-4c8f-a6e0-92f1b3c8d702
status: experimental
description: Detects a shell executing with euid 0 while the real uid is non-root, the unprivileged-to-root result of a Linux kernel local privilege escalation exploit (TL-2026-5445, T1068).
references:
  - https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/
  - https://www.bleepingcomputer.com/news/security/openais-new-gpt-6-cyber-ai-finds-81-firefox-flaws-948-linux-bugs/
author: AII-Detector — ThreadLinqs Intelligence
date: 2026-11-06
tags:
  - attack.privilege-escalation
  - attack.t1068
  - attack.t1548.001
logsource:
  product: linux
  service: auditd
detection:
  selection:
    type: 'SYSCALL'
    success: 'yes'
    euid: '0'
    exe|endswith:
      - '/sh'
      - '/bash'
      - '/dash'
      - '/zsh'
  filter_root:
    uid: '0'
  condition: selection and not filter_root
falsepositives:
  - Administrators deliberately running setuid-root shells for recovery tasks
  - Container runtimes with user namespace remapping
level: high

```

**Known false positives:**

- Administrators intentionally running setuid-root shells (bash -p) during recovery
- Container or user-namespace tooling that maps uid 0 inside a namespace

**Tuning notes:** Requires auditd execve/setuid syscall rules (-a always,exit -F arch=b64 -S execve). Exclude known service accounts via auid and namespace-mapped container hosts. Raise severity when the same auid ran a binary from /tmp or /dev/shm beforehand (see SPL-003).

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

- Vulnerable [indicator] Build Below 159 ([indicator] Exposure) (locked: unlock with TL Creds)
- Root Process Spawned by Non-Root Non-Broker Parent (Linux LPE) (locked: unlock with TL Creds)
- Non-Browser OpenAI API Access and Exploit-AI Tool Strings (Policy Hunt) (locked: unlock with TL Creds)
- [indicator] Spawning Shell or Download Utility (Linux) (locked: unlock with TL Creds)
- Setuid Bit Set on File in Writable Directory (Post-LPE Persistence) (locked: unlock with TL Creds)
- Command Line Referencing AI Exploit-Development Tooling (Policy Hunt) (locked: unlock with TL Creds)
- [indicator] Spawning Shell or LOLBin (Browser Exploitation, Windows Sysmon) (locked: unlock with TL Creds)
- Exploit Staging From Temp Path Followed by Root Shell (Linux LPE Chain) (locked: unlock with TL Creds)

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

## Sources

- [OpenAI's new GPT-6 Cyber AI finds 81 Firefox flaws, 948 Linux bugs (BleepingComputer)](https://www.bleepingcomputer.com/news/security/openais-new-gpt-6-cyber-ai-finds-81-firefox-flaws-948-linux-bugs/)
- [Mozilla Foundation Security Advisory 2026-118 (Firefox 159)](https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/)
- [OpenAI Expands Daybreak With Tiered Cyber Access](https://letsdatascience.com/news/openai-expands-daybreak-with-tiered-cyber-access-8d20cca5)
- [OpenAI's GPT-6 Astra reaches Critical cybersecurity capability tier](https://themodelwire.com/article/openais-gpt-6-astra-reaches-critical-cybersecurity-capability-tier-01M1MGH2HDND1VEB30AHVTW7YP)
- [OpenAI Daybreak for Frontline Defenders: $1 billion to protect critical infrastructure as GPT-6 Astra raises AI attack risk](https://www.kucoin.com/blog/openai-daybreak-for-frontline-defenders-1-billion-to-protect-critical-infrastructure-as-gpt-6-astra-raises-ai-attack-risk)
- [OpenAI Launches GPT-5.6-Cyber for Advanced Defensive Security Testing (eSecurity Planet)](https://www.esecurityplanet.com/security/news-openai-gpt-5-6-cyber-security-testing/)

## 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-2026-5445