Summary
New testing shows that OpenAI’s latest artificial intelligence, GPT-5.5, is just as capable at hacking as a specialized model that was recently restricted due to safety fears. The UK’s AI Security Institute (AISI) found that GPT-5.5 matches or even slightly beats Anthropic’s "Mythos Preview" in several cybersecurity challenges. While Anthropic limited access to its model to prevent it from being used for cyberattacks, OpenAI released GPT-5.5 to the public last week. These findings suggest that powerful hacking tools are now widely available, regardless of the restrictions placed on specific models.
Main Impact
The biggest impact of this news is the realization that general-purpose AI models have caught up to specialized "high-risk" models. For months, the tech industry debated whether certain AI models were too dangerous for the public to use. Anthropic argued that its Mythos model represented a major threat to digital security. However, the AISI data shows that GPT-5.5 can perform the same complex tasks. This means the "security wall" some companies tried to build around their AI may not be effective if other companies release similar tools to everyone.
Key Details
What Happened
The UK’s AI Security Institute put several top-tier AI models through a series of 95 different tests. These tests are known as "Capture the Flag" challenges. In these games, the AI must find hidden pieces of code, break into simulated websites, or solve complex math puzzles used in security. The goal was to see if the AI could act like a human hacker. GPT-5.5 was able to solve very difficult problems that required it to think through many steps without any help from a person.
Important Numbers and Facts
- Expert Level Success: GPT-5.5 passed 71.4% of the hardest tasks. Mythos Preview passed 68.6%.
- Speed and Cost: In one test, GPT-5.5 broke down a complex computer program in just over 10 minutes. The total cost for the computer power used was only $1.73.
- Network Attacks: In a test called "The Last Ones," which simulates a 32-step attack on a company network, GPT-5.5 succeeded 3 out of 10 times. Mythos succeeded 2 out of 10 times.
- Previous Limits: Before these two models, no AI had ever passed "The Last Ones" test even once.
- Physical Safety: Both models failed a test designed to see if they could shut down a power plant's cooling tower.
Background and Context
The UK government created the AI Security Institute to act as a watchdog. As AI gets smarter, experts worry it could be used to create new viruses or steal private data on a massive scale. Last month, Anthropic made headlines when it claimed its Mythos model was so good at hacking that it had to be kept under lock and key. They only allowed "critical industry partners" to use it. This created a lot of talk about whether AI was becoming too powerful. However, OpenAI’s release of GPT-5.5 shows that the technology is moving so fast that one company's "secret" power is quickly matched by others.
Public or Industry Reaction
The reaction to these tests has been a mix of surprise and concern. Many researchers believe that the "hype" around Mythos Preview was partly a marketing move by Anthropic to show how powerful their AI is. Now that GPT-5.5 has shown similar results, the focus is shifting toward how to defend against these tools. Security experts are pointing out that if a hack only costs $1.73 to perform, criminals could try thousands of attacks every day for a very low price. This makes the threat much more real for average businesses.
What This Means Going Forward
In the coming months, we will likely see a push for better digital defenses. Since AI can now handle multi-step attacks that involve planning and problem-solving, simple security measures may no longer be enough. Companies will need to use AI themselves to watch for these automated attacks. The good news is that AI still struggles with physical infrastructure. The failure of both models to disrupt the "Cooling Tower" simulation shows that our power grids and water systems are still somewhat safe from AI-driven sabotage for now. However, the gap is closing quickly.
Final Take
The latest data proves that high-level hacking skills are no longer a rare feature in AI. Whether a model is restricted or public, the capability exists. The focus must now move away from trying to hide these models and toward building stronger security systems that can withstand an AI-powered world. The era of cheap, fast, and automated hacking has arrived.
Frequently Asked Questions
Is GPT-5.5 better at hacking than Mythos Preview?
According to the UK AI Security Institute, GPT-5.5 performed slightly better on expert tasks, but the two models are very close in skill. Both are much more capable than any AI seen before 2026.
Can GPT-5.5 take down a power plant?
No. In the "Cooling Tower" simulation, GPT-5.5 failed to disrupt the control software. Current AI models still struggle with the complex systems used in physical infrastructure.
Why did Anthropic restrict access to its model?
Anthropic claimed that Mythos Preview posed a significant cybersecurity risk. They wanted to ensure that only trusted partners could use the model to prevent it from being used for harmful attacks.
