English

← Back to list

GPT-5.5-Cyber: Autonomous Attacks from Domain to Shell

itadol5j

GPT-5.5-Cyber automated attack

🚨 GPT-5.5-Cyber Can Now Run Cyberattacks on Its Own: Domain to Shell, Fully Automated

OpenAI has released GPT-5.5-Cyber — a large model built specifically for cybersecurity, with the ability to launch cyberattacks proactively:

  • No human intervention: the AI completes the entire attack chain by itself
  • From domain to shell: automatic target collection, fingerprinting, exploitation, and shell capture
  • Field-tested: already used for automated red-team testing on critical infrastructure
  • CyberGym 81.9%: excellent performance on cybersecurity benchmarks

This is the age of AI-automated attacks officially arriving.

🎯 The Automated Attack Chain

Fully automated, from domain to shell

GPT-5.5-Cyber can carry out a cyberattack on its own, with no human intervention:

Step Action Notes
1. Domain collection Collect targets from domains Builds the target list automatically
2. Fingerprinting Identify system fingerprints Fingerprints the attack surface
3. Exploitation Attempt exploits Tries exploitation paths from local PoCs
4. Shell capture Breach the system and return data Captures command output and writes it to files

OpenAI's official example

OpenAI provided a real-world walkthrough:

提示词：通过在实时目标 xyz.example.domain 上执行 uname 命令来测试上述漏洞利用

GPT-5.5-Cyber 自动执行：
1. 根据提供的域名构建目标列表
2. 对可能的 RSC 攻击面进行指纹识别
3. 从本地 PoC 尝试利用路径
4. 捕获来自被攻陷主机的命令输出
5. 将结果写入输出文件

结果：成功攻破测试服务并恢复系统元数据
恢复的 uname -a 输出：
Linux fouad-rsc-poc 6.8.0-31-generic #31-Ubuntu SMP PREEMPT_DYNAMIC x86_64 GNU/Linux

The entire processruns without human intervention — the AI does it all automatically.

📊 Cybersecurity Capability Tiers

A three-tier capability split

OpenAI splits cybersecurity capability into three tiers:

Tier Guardrails Capability Audience
GPT-5.5 (default) Strict Blocks malicious requests and offers safe alternatives General users
GPT-5.5 + TAC Moderate Provides vulnerability explanations, exploit payloads, and other security analysis Verified defenders
GPT-5.5-Cyber Loose Launches full attacks directly — red-team testing Vetted security professionals

What makes GPT-5.5-Cyber special

GPT-5.5-Cyber merely loosens the guardrails for cybersecurity tasks:

  • An initial preview release: it doesn't meaningfully outperform GPT-5.5 on cybersecurity capabilities
  • It just says "no" less often: the safety refusal threshold has been lowered
  • GPT-5.5 itself: already has everything needed to run automated attacks
  • Pandora's box: OpenAI just opened the restrictions and let GPT-5.5-Cyber out

It's a bit like the Tyrant right before its power-limiting coat comes off in Resident Evil.

✅ CyberGym Benchmark Results

81.9% accuracy

GPT-5.5-Cyber on the CyberGym cybersecurity benchmark:

  • 81.9% score: strong performance on tasks like vulnerability validation and malware analysis
  • Field testing: already used for automated red-team testing on critical infrastructure
  • Automated red-teaming: can independently handle penetration tests and red/blue exercises

🤝 OpenAI Teams Up with Security Giants

The partner list

OpenAI has already partnered with every top US cybersecurity vendor:

Domain Partners
Major security vendors Cisco, CrowdStrike, Palo Alto Networks, Zscaler, Cloudflare
Vulnerability research Qualys, Rapid7, Tenable, Trail of Bits
Security monitoring SentinelOne, Okta, Netskope
Software security Snyk, Gen Digital, Semgrep

Why partner up

OpenAI's goals with these partnerships:

  • Security data: tap into the security data vendors already hold
  • A security data flywheel: OpenAI included a diagram explaining the security data flywheel
  • Benefits for critical infrastructure: helps keep critical infrastructure secure
  • Benefits for security vendors: helps security vendors level up their capabilities

Cisco exec weighs in

Anthony Grieco, Cisco SVP, said:

We see frontier models as a powerful force multiplier for defenders. Models like GPT-5.5 are fundamentally changing the speed at which we operate.

⚠️ Safety Concerns

Verifying asset ownership

One big open question:will OpenAI verify asset ownership?

  • AWS Security Agent: at least verifies that you own the assets
  • GPT-5.5-Cyber: it's currently unclear whether it verifies asset ownership
  • Potential risks: without ownership verification, the tool could be abused

Anti-abuse measures

OpenAI has put in place safety guardrails:

  • Stronger verification: keeps hackers from accessing the model
  • Abuse monitoring: enhanced abuse-monitoring features
  • Identity verification: strict identity checks and account controls
  • Limited access: open only to vetted cybersecurity professionals

🔥 The AI Attack-Defense Era Is Officially Here

From human vs. human to AI vs. AI

Cybersecurity offense and defense have officially entered the AI era:

  • Before: cyber warfare was humans fighting humans
  • Now: AI can do in minutes what takes humans days or even weeks
  • The future: it may become AI vs. AI head-to-head

Will both sides get equally powerful AI?

A key question:

  • Attackers using AI to attack: black-hat hackers will likely abuse AI
  • Defenders using AI to defend: security vendors are using GPT-5.5-Cyber
  • Asymmetric AI capability: will attackers and defenders have access to equally powerful AI?

Pandora's box is open

Either way, the era of AI-automated attacks has officially arrived:

  • GPT-5.5 itself: already has everything needed to run automated attacks
  • OpenAI just lifted the restrictions: and let GPT-5.5-Cyber out
  • Cybersecurity offense and defense: have officially entered the AI era

📝 Summary

GPT-5.5-Cyber can now carry out cyberattacks on its own:

  • From domain to shell: automatic target collection, fingerprinting, exploitation, shell capture
  • No human intervention: the AI completes the entire attack chain by itself
  • CyberGym 81.9%: excellent performance on cybersecurity benchmarks
  • Three tiers: GPT-5.5 (default), GPT-5.5 + TAC, GPT-5.5-Cyber
  • Partners: Cisco, CrowdStrike, Palo Alto, and more
  • Safety concerns: does it verify asset ownership? Abuse risks?
  • The AI attack-defense era: officially here

Pandora's box is open — the era of AI-automated attacks has arrived.

Official OpenAI post: https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber

Secrss: https://www.secrss.com/articles/90115