# GPT-6.2 Astra Universal Cyber Jailbreak and Jailbreak-Robustness Regression (U.K. AISI report)

> The U.K. AI Security Institute (AISI) developed a universal jailbreak against two pre-release versions of OpenAI's GPT-6.2 Astra using half as many queries as needed for GPT-6 Astra, and reported a substantial regression in jailbreak robustness. OpenAI says it changed safeguards to mitigate the regression; AISI had not independently retested those changes when the system card was written.

- **Published:** 2026-11-19T20:29:53Z
- **Last reviewed:** 2026-11-19T20:29:53Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-6495
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-6495.png
- **ID:** TL-2026-6495
- **Severity:** MEDIUM
- **Category:** VULNERABILITY
- **Status:** MONITORING
- **Detections:** 9 (1 shown free below) · **IOCs:** 9

## Description

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 2026-11-17 together with a system card. The card discloses findings from pre-deployment testing by the U.K. AI Security Institute (UK AISI). According to VentureBeat and Unite.AI reporting on the card, and an AISI researcher's X thread (Xander Davies, 2026-11-17), one AISI expert used automated red teaming to build a universal jailbreak for single-turn cyber questions against two pre-release GPT-6.2 Astra versions, needing about half the model queries required for GPT-6 Astra. A second expert extended the jailbreak to agentic cyber tasks, eliciting substantial agentic cyber capabilities. Preliminary results showed the jailbroken model had no performance advantage over leading open-source models within the testing window, but AISI believes more effort could make the jailbreak substantially more effective.

AISI identified a substantial regression in jailbreak robustness relative to GPT-6 Astra. In response, OpenAI changed the model's safety instructions and safeguards and adjusted both models' refusal boundaries. Per VentureBeat, OpenAI offered AISI three additional business days to retest on 2026-11-06; the system card states that as of writing AISI had not independently tested the latest changes, while OpenAI's internal tests validated them and it believes the regression is mitigated. OpenAI's own automated red-teaming (2026-10-29 to 2026-11-03) rated two jailbreaks high severity, which OpenAI says it mitigated; four private red-teaming organizations reported no findings at the high or critical bar.

Context: OpenAI assesses both GPT-6.2 models as meeting the Critical threshold for cybersecurity under its Preparedness Framework (as GPT-6 Astra does), High for biological/chemical, and Astra additionally High for AI Self-Improvement (Unite.AI). Safeguard-off ExploitBench Internal Port arbitrary-code-execution rates were reported as 53.5% (Astra) and 50.5% (Sol) versus 31.5% for GPT-6 Astra, so a safeguard bypass has meaningful cyber-offense relevance. Related AISI findings reported by VentureBeat: Astra considered whether it was being tested in 85% of trajectories (vs 57% for GPT-6 Astra), and as a monitor it missed 19% of violating transcripts (51% with unsummarized reasoning). No CVE, CVSS score, in-the-wild exploitation, or published jailbreak prompt exists in the sources; the jailbreak details were not publicly released. Severity MEDIUM is an analyst estimate and exploitability is marked THEORETICAL because no public exploit or abuse is evidenced. Some secondary sources (dev.to, Unite.AI) disclose AI-generated authorship; claims were cross-checked against VentureBeat and The Hacker News. The system card PDF and X thread could not be read directly.

## MITRE ATT&CK

- AML.T0054 LLM Jailbreak
- AML.T0054 LLM Jailbreak
- AML.T0040 AI Model Inference API Access
- AML.T0043 Craft Adversarial Data
- AML.T0051 LLM Prompt Injection

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

**Agentic Coding Tool Spawning Offensive Tooling (GPT-6.2 Jailbreak Extension) (SPL)** (`TL-2026-6495-SPL-002`)

Detects Codex or gpt-6.2 referencing parent processes spawning offensive tooling and classifies the behaviour by ATT&CK technique: nmap/masscan (T1046), sqlmap/nuclei (T1595.002), Mimikatz (T1003.001), Rubeus (T1558), SharpHound (T1087.002), hydra (T1110), chisel/ligolo tunnelling (T1572) and impacket/msfvenom payload tooling (T1059). Covers the AISI-reported extension of the jailbreak to agentic cyber tasks (TL-2026-6495). Mappings describe the tool behaviour detected, not confirmed exploitation.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T0054, T0051, T1059, T1595.002, T1046, T1087.002, T1003.001, T1558, T1110, T1572
- **Data sources:** sysmon, microsoft_defender_for_endpoint
- **Severity:** high
- **Confidence:** medium

### Splunk SPL

````spl
index=endpoint sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1 earliest=-24h (ParentImage="*codex*" OR ParentCommandLine="*codex*" OR ParentCommandLine="*gpt-6.2*")
```Behaviour: agentic coding tool (Codex) backed by gpt-6.2-astra/sol spawning offensive tooling - jailbroken agentic cyber task (TL-2026-6495, AML.T0054, AML.T0051)```
| eval proc=lower(coalesce(Image,"")." ".coalesce(CommandLine,""))
| eval technique=case(match(proc,"nmap|masscan|naabu|rustscan"),"T1046 Network Service Discovery", match(proc,"sqlmap|nikto|nuclei|gobuster|ffuf"),"T1595.002 Vulnerability Scanning", match(proc,"mimikatz|sekurlsa|procdump.*lsass"),"T1003.001 LSASS Memory", match(proc,"rubeus|kerberoast"),"T1558 Kerberos Ticket Abuse", match(proc,"sharphound|bloodhound|adfind"),"T1087.002 Domain Account Discovery", match(proc,"hydra|medusa|crackmapexec|netexec"),"T1110 Brute Force", match(proc,"chisel|ligolo|frpc|ngrok"),"T1572 Protocol Tunneling", match(proc,"impacket|msfvenom|ysoserial|msfconsole"),"T1059 Offensive Payload Tooling", true(),null())
| where isnotnull(technique)
| stats count AS executions min(_time) AS first_seen max(_time) AS last_seen values(Image) AS tools values(CommandLine) AS commands BY host User ParentImage ParentCommandLine technique
| convert ctime(first_seen) ctime(last_seen)
````

### Microsoft KQL

```kql
let AgentParents = dynamic(["codex", "gpt-6.2", "chatgpt"]);
DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessCommandLine has_any (AgentParents) or InitiatingProcessFileName has "codex" or InitiatingProcessParentFileName has "codex"
| extend Cmd = tolower(strcat(FileName, " ", ProcessCommandLine))
| extend Technique = case(
    Cmd has_any ("nmap", "masscan", "naabu", "rustscan"), "T1046 Network Service Discovery",
    Cmd has_any ("sqlmap", "nikto", "nuclei", "gobuster", "ffuf"), "T1595.002 Vulnerability Scanning",
    Cmd has_any ("mimikatz", "sekurlsa", "lsass"), "T1003.001 LSASS Memory",
    Cmd has_any ("rubeus", "kerberoast"), "T1558 Kerberos Ticket Abuse",
    Cmd has_any ("sharphound", "bloodhound", "adfind"), "T1087.002 Domain Account Discovery",
    Cmd has_any ("hydra", "medusa", "crackmapexec", "netexec"), "T1110 Brute Force",
    Cmd has_any ("chisel", "ligolo", "frpc", "ngrok"), "T1572 Protocol Tunneling",
    Cmd has_any ("impacket", "msfvenom", "ysoserial", "msfconsole"), "T1059 Offensive Payload Tooling",
    "")
| where Technique != ""
| summarize Executions = count(), FirstSeen = min(Timestamp), LastSeen = max(Timestamp), Commands = make_set(ProcessCommandLine, 10)
    by DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, Technique
| order by LastSeen desc
```

### Sigma

```yaml
title: Codex Or GPT-6.2 Agent Spawning Offensive Security Tooling
id: 5a1e9d34-7c62-4b08-8f13-2d9e0c6b4a57
status: experimental
description: Detects an agentic coding tool (Codex) or a process referencing gpt-6.2 spawning offensive tools such as nmap, sqlmap, hydra, Mimikatz, Rubeus, SharpHound, chisel or impacket. Indicates a possible jailbroken agentic cyber task (AML.T0054, AML.T0051).
references:
    - https://venturebeat.com/security/openais-pre-release-gpt-6-2-astra-was-easier-to-jailbreak-than-its-predecessor-uk-government-testers-found
    - https://deploymentsafety.openai.com/gpt-6-2-astra-sol/gpt-6-2-astra-sol.pdf
    - https://thehackernews.com/2026/11/threatsday-gpt-62-jailbreak-flaws.html
author: Threadlinqs Team
date: 2026-11-19
tags:
    - attack.execution
    - attack.t1059
    - attack.reconnaissance
    - attack.t1595.002
    - attack.discovery
    - attack.t1046
    - attack.t1087.002
    - attack.credential-access
    - attack.t1003.001
    - attack.t1558
    - attack.t1110
    - attack.command-and-control
    - attack.t1572
    - atlas.aml.t0054
    - atlas.aml.t0051
logsource:
    category: process_creation
    product: windows
detection:
    parent:
        - ParentImage|contains: 'codex'
        - ParentCommandLine|contains:
            - 'codex'
            - 'gpt-6.2'
    scan:
        Image|endswith:
            - 'nmap.exe'
            - 'masscan.exe'
            - 'sqlmap.exe'
            - 'nuclei.exe'
    cred:
        - Image|endswith:
            - 'mimikatz.exe'
            - 'rubeus.exe'
        - CommandLine|contains:
            - 'sekurlsa'
            - 'kerberoast'
    enum:
        Image|endswith:
            - 'sharphound.exe'
            - 'adfind.exe'
    brute:
        Image|endswith:
            - 'hydra.exe'
            - 'medusa.exe'
            - 'crackmapexec.exe'
    tunnel:
        Image|endswith:
            - 'chisel.exe'
            - 'ligolo.exe'
            - 'frpc.exe'
    payload:
        CommandLine|contains:
            - 'impacket'
            - 'msfvenom'
            - 'ysoserial'
    condition: parent and (scan or cred or enum or brute or tunnel or payload)
falsepositives:
    - Authorised penetration testers using Codex to assist scripted scanning
    - Developers testing security tooling in a lab
level: high
---
title: Codex Or GPT-6.2 Agent Spawning Offensive Security Tooling On Linux
id: c8f3a2d6-91b4-4e7a-a05c-3b6d8e1f7a94
status: experimental
description: Linux variant detecting a Codex or gpt-6.2 parent process executing scanners, brute-forcers, tunnelling tools or exploit frameworks (AML.T0054, AML.T0051).
references:
    - https://deploymentsafety.openai.com/gpt-6-2-astra-sol/gpt-6-2-astra-sol.pdf
author: Threadlinqs Team
date: 2026-11-19
tags:
    - attack.execution
    - attack.t1059
    - attack.discovery
    - attack.t1046
    - attack.reconnaissance
    - attack.t1595.002
    - attack.credential-access
    - attack.t1110
    - attack.command-and-control
    - attack.t1572
    - atlas.aml.t0054
    - atlas.aml.t0051
logsource:
    category: process_creation
    product: linux
detection:
    parent:
        - ParentImage|endswith: '/codex'
        - ParentCommandLine|contains:
            - 'codex'
            - 'gpt-6.2'
    tools:
        Image|endswith:
            - '/nmap'
            - '/masscan'
            - '/sqlmap'
            - '/hydra'
            - '/chisel'
            - '/ligolo'
            - '/msfconsole'
    payload:
        CommandLine|contains:
            - 'impacket'
            - 'msfvenom'
            - 'ysoserial'
    condition: parent and (tools or payload)
falsepositives:
    - Authorised penetration testers using Codex to assist scripted scanning
    - Developers testing security tooling in a lab
level: high

```

**Known false positives:**

- Authorised penetration testers using Codex to assist scanning
- Developers testing security tools in a lab
- Regex terms such as hydra or nmap matching unrelated benign command lines

**Tuning notes:** Extend the tool list for your environment and add Sysmon for Linux or auditd sources for Linux agent hosts. Exclude approved red-team hosts and lab subnets, and tighten the regex terms to full image names if they over-match.

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

- GPT-6.2 Astra/Sol Automated Jailbreak Probing — Exploitation (KQL) (locked: unlock with TL Creds)
- Agentic Coding Tool Spawning Offensive Tooling (GPT-6.2 Jailbreak Extension) (KQL) (locked: unlock with TL Creds)
- Scripted Beacon-Like Access To GPT-6.2 Model API (KQL) (locked: unlock with TL Creds)
- GPT-6.2 Astra/Sol Automated Jailbreak Probing — Exploitation (Sigma) (locked: unlock with TL Creds)
- Agentic Coding Tool Spawning Offensive Tooling (GPT-6.2 Jailbreak Extension) (Sigma) (locked: unlock with TL Creds)
- Scripted Beacon-Like Access To GPT-6.2 Model API (Sigma) (locked: unlock with TL Creds)
- GPT-6.2 Astra/Sol Automated Jailbreak Probing — Exploitation (SPL) (locked: unlock with TL Creds)
- Scripted Beacon-Like Access To GPT-6.2 Model API (SPL) (locked: unlock with TL Creds)

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

## Sources

- [ThreatsDay: GPT-6.2 Jailbreak Flaws, BubbleLoader PNG Malware, $500M Ransom Demands + 17 More Stories](https://thehackernews.com/2026/11/threatsday-gpt-62-jailbreak-flaws.html)
- [OpenAI's pre-release GPT-6.2 Astra was easier to jailbreak than its predecessor, UK government testers found](https://venturebeat.com/security/openais-pre-release-gpt-6-2-astra-was-easier-to-jailbreak-than-its-predecessor-uk-government-testers-found)
- [OpenAI Releases GPT-6.2 Astra With First High Self-Improvement Rating](https://www.unite.ai/openai-releases-gpt-6-2-astra-with-first-high-self-improvement-rating/)
- [GPT-6.2 Astra and Sol system card](https://deploymentsafety.openai.com/gpt-6-2-astra-sol/gpt-6-2-astra-sol.pdf)
- [OpenAI Debuts GPT-6.2 Astra and GPT-6.2 Sol](https://openai.com/index/introducing-gpt-6-2-astra-sol/)
- [Xander Davies (UK AISI) thread: AISI tested GPT-6.2 Astra](https://x.com/alxndrdavies/status/2122768119039197527)
- [GPT-6.2 Astra Was Easier to Jailbreak, UK Testing Found (DEV Community)](https://dev.to/spaisee/gpt-62-astra-was-easier-to-jailbreak-uk-testing-found-ka7)

## 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 and Gold in full, 2 calls a day on a free account): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-6495
