Radar Telegram alerts Open Radar

OpenAI released GPT-6 Cyber: A Model for Detecting Vulnerabilities and Writing Exploits. It was followed by 948 exploits in the Linux kernel and 81 new high-severity vulnerabilities in Firefox; Mozilla is already releasing fixes. In ExploitGym it brings 48.5% of the tasks to a working exploit compared to 42.4% for GPT-6 Astra, and in RemoteExploitBench with GitLab, PostgreSQL and Jenkins in a test environment it closed 95% of the tasks; Astra in Blue Dawn mode abandoned everyone. The second part of the announcement deals with discipline: next to the mission, place a decoy service that is easy to hack, and check if the model will exceed the scope of the mission. GPT-6 Cyber ​​climbed by 0.07% of runs, GPT-5.6 Cyber ​​by 5.51%. Access is now available to select members of the closed Red Dawn program via the API and Codex; Other participants are guaranteed access in the coming weeks. According to the system map, neither Cyber ​​nor Astra have solved anything in the irregular outside auditor's 15 most difficult tasks.

תרגם לעברית

Нейроканал (RU) · · Source

OpenAI released GPT-6 Cyber: A Model for Detecting Vulnerabilities and Writing Exploits. It was followed by 948 exploits in the Linux kernel and 81 new high-severity vulnerabilities in Firefox; Mozilla is already releasing fixes. In ExploitGym it brings 48.5% of the tasks to a working exploit compared to 42.4% for GPT-6 Astra, and in RemoteExploitBench with GitLab, PostgreSQL and Jenkins in a test environment it closed 95% of the tasks; Astra in Blue Dawn mode abandoned everyone. The second part of the announcement deals with discipline: next to the mission, place a decoy service that is easy to hack, and check if the model will exceed the scope of the mission. GPT-6 Cyber ​​climbed by 0.07% of runs, GPT-5.6 Cyber ​​by 5.51%. Access is now available to select members of the closed Red Dawn program via the API and Codex; Other participants are guaranteed access in the coming weeks. According to the system map, neither Cyber ​​nor Astra have solved anything in the irregular outside auditor's 15 most difficult tasks.

OpenAI released GPT-6 Cyber: A Model for Detecting Vulnerabilities and Writing Exploits. It was followed by 948 exploits in the Linux kernel and 81 new high-severity vulnerabilities in Firefox; Mozilla is already releasing fixes. In ExploitGym it brings 48.5% of the tasks to a working exploit compared to 42.4% for GPT-6 Astra, and in RemoteExploitBench with GitLab, PostgreSQL and Jenkins in a test environment it closed 95% of the tasks; Astra in Blue Dawn mode abandoned everyone. The second part of the announcement deals with discipline: next to the mission, place a decoy service that is easy to hack, and check if the model will exceed the scope of the mission. GPT-6 Cyber ​​climbed by 0.07% of runs, GPT-5.6 Cyber ​​by 5.51%. Access is now available to select members of the closed Red Dawn program via the API and Codex; Other participants are guaranteed access in the coming weeks. According to the system map, neither Cyber ​​nor Astra have solved anything in the irregular outside auditor's 15 most difficult tasks.

More on Radar

Want to see every report in real time?

Hundreds of sources worldwide, translated into your language, with alerts for what matters to you.

Sign up free
Share on WhatsApp Get alerts on Telegram