OpenAI’s new artificial intelligence model, GPT-5.5, has the capacity to find security flaws at a level equivalent to the Mythos model, which competitor Anthropic chose to keep restricted to the public, according to the president of the creator of ChatGPT, Greg Brockman.
“We believe in empowering people with AI,” said Brockman at a press conference this Thursday (23). The model is now available on ChatGPT and Codex, the company’s programming and automation platform. The company should add it to the API (way of calling AI via programming code) in the coming days.
According to the executive, there was gradual work over two years that allowed the technology to be launched safely.
The company works with independent experts to identify risks and impose barriers to artificial intelligence in sensitive areas such as biological weapons and cybersecurity – a useful skill for companies to reinforce their systems and for criminals to find loopholes.
The general public will have access to a version with limited capabilities in activities related to cybersecurity, while companies in the sector can request access to an unlimited version of OpenAI technology from February 8th.
While Mythos scored 83.1% on a cybersecurity skill test, GPT-5.5 scored 81.8% (the highest score recorded by a widely available model).
“Mythos is a permissive model in terms of cybersecurity. The model we are launching has mitigations, it is not a permissive model,” OpenAI said.
ANTHROPIC COMPETITION
Claude’s creator, Anthropic, announced on the 8th that it had developed a “system too powerful to be released to the public”. At the time, it guaranteed access to technology to only a select group of 40 large American companies, including big techs and banks.
Logan Graham, head of a team at Anthropic that tests new models for dangerous capabilities, called the new model “the starting point for what we believe will be a game-changer in the industry.”
Brazilian Mike Krieger, an executive who leads Anthropic’s cutting-edge laboratory, said the artificial intelligence startup realized the risks of a recently announced new advanced model and chose to “take a step back to prepare the world for this challenge.”
Since the announcement of OpenAI’s main competitor, policymakers have warned of the risk that advanced AI poses to cybersecurity. On the one hand, the president of the European Central Bank, Christine Lagarde, praised Anthropic’s thrift in keeping technology away from criminals. On the other hand, representatives of Chinese diplomacy criticized the monopoly on technical advancement in the hands of American companies.
The OpenAI program also has limited access only to partners released by the creator of ChatGPT. The company’s terms of use also limit access to the technology to countries under US government restrictions, expanded during Trump’s administration. Interested Brazilian businesses can request participation in the program via this link.
According to the engineer responsible for GPT-5.5 and the company’s current vice president of research, Amelia Glaese, OpenAI assesses that the risks of AI advance gradually. Therefore, the company has carried out repeated tests in recent months
“There are three principles by which we define our approach: trial and error development, resilient ecosystem and democratic access,” Glaese said at the press conference.
TECHNICAL ADVANCES
According to Brockman, the new OpenAI model did not overcome the barrier of 10 trillion parameters, a measure of power associated with AI models. Leaks from Anthropic, which had its system hacked this month, indicate that Mythos would have this scale of magnitude.
OpenAI’s chief scientist, Jakub Pachocki, states that the advancement of technology, since 2024, has been slower in terms of scale – between 2020 and 2025, models jumped from millions of parameters to more than 1 trillion. The main changes, he says, are linked to the chain of reasoning, a technique in which AI draws up action plans and corrects itself until it finds a response it considers appropriate.
Also according to Pachocki, GPT-5.5 is the first model that will present your action plan for a task before executing it. This way, the user can guide the approach before wasting time and resources in the wrong way.
FRICTIONS IN WASHINGTON
Since the beginning of the year, Anthropic and OpenAI have been at the center of an internal dispute within the Trump administration’s Department of Defense. The creator of Claude refused to provide technology that would be used in the development of autonomous weapons and the surveillance of American citizens. For this reason, she was punished with restrictions and breach of contracts.
The ChatGPT developer took over its competitor’s contracts, saying the US government agreed to its ethical limits on security and privacy.
Source: www.noticiasaominuto.com.br
Source link
