OpenAI just showed five state-affiliated hacking crews the door. In a joint operation with Microsoft Threat Intelligence, the company terminated accounts linked to threat actors from China, Iran, North Korea, and Russia — groups with names like Charcoal Typhoon, Crimson Sandstorm, and Forest Blizzard. It’s a rare public peek into how nation-states are poking at AI models, and what they’re actually getting out of them.
The reality is less dramatic than the sci-fi warnings suggest. These groups weren’t building autonomous cyberweapons or cracking unbreakable encryption. Charcoal Typhoon (China) used ChatGPT to debug scripts and draft phishing content. Salmon Typhoon (China) translated technical papers and researched intelligence agencies. Iran’s Crimson Sandstorm generated spear-phishing material and looked into malware evasion techniques. North Korea’s Emerald Sleet scoped out defense experts in the Asia-Pacific and worked on basic scripting. Russia’s Forest Blizzard mostly hunted for open-source intel on satellite communications and radar imaging.
OpenAI’s own red teaming, done alongside external cybersecurity experts, backs up the assessment: GPT-4 offers “only limited, incremental capabilities for malicious cybersecurity tasks” — nothing you can’t already accomplish with publicly available, non-AI tools. That doesn’t mean there’s no risk, but it does puncture the breathless narrative that LLMs are a cheat code for cyberattacks. The real value for these actors, at least for now, appears to be mundane productivity: translation, research, fixing sloppy code.
The company’s response goes beyond just flipping a kill switch on accounts. OpenAI is leaning on its Intelligence and Investigations team to use its own models to hunt adversaries, analyze how they interact with the platform, and assess their broader intentions. They’re also sharing intel with industry partners and baking lessons from real-world misuse back into safety mitigations. The thinking is straightforward: watch what the most sophisticated attackers try today, because that’s what less-skilled operators will attempt tomorrow.
