← Back to the post

Verification log

OpenAI says GPT-6 Cyber goes out of bounds less than 1% of the time

Checked

86

Confirmed

82

Re-cited

3

Removed

1

Written by hand-written and layer-2 checked by Claude (Opus 5.5). "Re-cited" means the fact was found in another source about the same story, and the citation was corrected. Removed statements never reached the post.

• Confirmed headline

  OpenAI says GPT-6 Cyber goes out of bounds less than 1% of the time

• Confirmed quote

  It’s over 99 percent of the time doing what you ask.

• Confirmed dek

  OpenAI ships a safer cybersecurity model and will buy a code-tools maker, Claude gets read-only access to bank accounts, Meta warns that voters are turning on data centres, and Google makes AI video cheaper.

• Confirmed what_happened[0]

  OpenAI says its new GPT-6 Cyber model went out of bounds less than 1 percent of the time in testing, and it is buying Pierre.

• Confirmed what_happened[1]

  Anthropic's Claude can now connect to US users' bank accounts through Plaid, with read-only access and up to two years of transactions.

• Confirmed what_happened[2]

  Meta's president says the midterm results will make data centres harder to build, as people tie them to how they feel about AI.

• Confirmed sections[0]

  OpenAI is launching GPT-6 Cyber, the latest version of its cybersecurity AI model.

• Confirmed sections[0]

  OpenAI says it is better at finding vulnerabilities and less likely to go rogue and hack other platforms in the process.

• Confirmed sections[0]

  During testing, GPT-6 Cyber went out of bounds less than 1 percent of the time, down from around 5 percent with GPT-5.6 Cyber, according to OpenAI.

• Confirmed sections[0]

  The test, called ExploitGym HoneyPot, is modeled after the Hugging Face Incident, in which OpenAI's AI agents breached the open-source AI community to find answers to a problem they were tasked with solving.

• Confirmed sections[0]

  Fouad Matin, the head of OpenAI's cyber preparedness and safety team, said the model does what it is asked over 99 percent of the time.

• Confirmed sections[0]

  TechCrunch reports that Max Justicz, co-founder of the startup Charge Robotics, used an OpenAI model to find a severe vulnerability that could let hackers take control of servers running Ubuntu.

• Confirmed sections[0]

  He reported it to the Debian Project, which issued a patch, and there has been no known evidence of exploitation.

• Confirmed sections[0]

  Separately, OpenAI announced that it will acquire Pierre Computer Company, the company behind Code Storage and a family of open-source developer tools.

• Confirmed sections[0]

  Pierre's Trees and Diffs libraries already help users browse files and review changes in the Codex app.

• Confirmed sections[0]

  OpenAI says Pierre's open-source projects will remain open-source, and customers can continue using Code Storage after the acquisition closes.

• Confirmed section title

  OpenAI ships GPT-6 Cyber and will buy a code-tools maker

• Confirmed card[0] hook

  GPT-6 Cyber went out of bounds less than 1% of the time in testing.

• Confirmed card[0] point

  OpenAI says it finds vulnerabilities better and is less likely to hack other platforms.

• Confirmed card[0] point

  The older GPT-5.6 Cyber went out of bounds around 5 percent of the time.

• Confirmed card[0] point

  OpenAI is also buying Pierre, maker of code tools used in the Codex app.

• Confirmed card[0] take

  If you look after software, AI bug hunting is becoming safer to try.

• Re-cited card[0] ask

  Would you trust an AI to hunt bugs in your code?

• Confirmed card[0] duo line 1

  OpenAI made a new AI that hunts for security holes in software.

• Confirmed card[0] duo line 3

  OpenAI says the old version went out of bounds around 5 percent of the time.

• Confirmed card[0] duo line 5

  The new one does it less than 1 percent of the time, OpenAI says.

• Confirmed card[0] duo line 7

  One startup founder used an OpenAI model to find a serious bug that put Ubuntu servers at risk.

• Confirmed sections[1]

  Anthropic announced on Thursday that Claude is getting a new tool called Personal Finance, designed to help users keep track of their money.

• Confirmed sections[1]

  Users can connect their bank and credit card accounts through Plaid and ask Claude questions about their finances.

• Confirmed sections[1]

  Anthropic says access is read-only, meaning Personal Finance can't move, manage, or invest users' money.

• Confirmed sections[1]

  Starting this week, Pro, Max and Ultra subscribers in the US can connect their financial accounts, Engadget reports.

• Confirmed sections[1]

  Plaid's backend has connections to more than 12,000 financial institutions, including Citi and Chase.

• Confirmed sections[1]

  Anthropic says the tool can flag potential issues such as duplicate charges, unexpected fees and steadily increasing bills.

• Confirmed sections[1]

  Claude's access covers balances, credit limits and up to two years of transaction history, and conversations in the Personal Finance hub aren't used to train its models.

• Confirmed sections[1]

  With the launch, Claude is playing catch-up to OpenAI's ChatGPT, which introduced similar personal finance tools in May, according to TechCrunch.

• Confirmed sections[1]

  Anthropic says Personal Finance has no ads or sponsored recommendations, and it doesn't offer credit cards, loans, or savings accounts of its own.

• Confirmed sections[1]

  Engadget notes that Anthropic didn't design the tool to give people investment advice.

• Confirmed section title

  Claude gets a read-only view of your bank accounts

• Confirmed card[1] hook

  Claude can now read your bank accounts, but it can't move money.

• Confirmed card[1] point

  Users connect bank and card accounts through Plaid and ask about their money.

• Confirmed card[1] point

  It can flag duplicate charges, unexpected fees and steadily increasing bills.

• Confirmed card[1] point

  Access is read-only and covers up to two years of transactions.

• Confirmed card[1] take

  Check exactly what an AI can see before you connect your accounts.

• Re-cited card[1] ask

  Would you let an AI read your bank account?

• Confirmed card[1] duo line 1

  Claude can now look at your bank and credit card accounts.

• Confirmed card[1] duo line 3

  You ask where your money went, and it sorts your spending for you.

• Confirmed card[1] duo line 5

  It can spot duplicate charges, unexpected fees and bills that keep going up.

• Confirmed card[1] duo line 7

  And it can only look. It can't move, manage or invest your money.

• Confirmed sections[2]

  Meta president Dina Powell McCormick said the Democratic wins in the midterm elections are going to make it a lot harder for companies to build data centres.

• Confirmed sections[2]

  Speaking at the WSJ Tech conference on Wednesday, she said the industry has not done as good a job as it needed to explain the benefits of these investments.

• Confirmed sections[2]

  She said the backlash against data centres, which was on the ballot during this election season, is also tied to how people feel about AI technology overall.

• Confirmed sections[2]

  Meta has data centres in 23 states, she said, and five of them, including one in Texas, are planned as projects of 1 gigawatt or more.

• Confirmed sections[2]

  TechCrunch reports that Meta is now worried it could lose some of its investments after politicians ran on anti-data-center platforms during the midterms.

• Confirmed sections[2]

  Powell McCormick said communities only want data centres when they benefit in some way.

• Confirmed sections[2]

  In Meta's successful projects, she said, the company pays for its own electricity and directs tax surpluses from the data centres toward teachers.

• Confirmed sections[2]

  Meta believes it could turn around negative feelings about AI if it can show the value of both the data centres and the technology itself.

• Confirmed section title

  Meta says election backlash will make data centres harder to build

• Confirmed card[2] hook

  Meta says the midterm results will make data centres harder to build.

• Confirmed card[2] point

  Meta's president says people link data centres to how they feel about AI.

• Confirmed card[2] point

  Meta has data centres in 23 states, five planned at 1 gigawatt or more.

• Confirmed card[2] point

  Meta worries it could lose investments after anti-data-center campaigns.

• Confirmed card[2] take

  The AI race now runs through town halls as well as labs.

• Re-cited card[2] ask

  Would you want a data centre in your town?

• Confirmed card[2] duo line 1

  Meta says the midterm election results will make data centres much harder to build.

• Confirmed card[2] duo line 3

  Its president says people see a data centre and think about AI in general.

• Confirmed card[2] duo line 5

  Meta has data centres in 23 states, and five are planned as huge power projects.

• Confirmed card[2] duo line 7

  Meta says towns only want them when they get something back, so it pays its own electricity bill.

• Confirmed sections[3]

  Google DeepMind is rolling out Gemini Omni 1.2 Flash, which offers native 15-second video creation, up from 10 seconds with Omni 1.1.

• Confirmed sections[3]

  Google says the update also brings improved reference adherence and multi-turn conversational editing, so users can adjust a clip without starting over.

• Confirmed sections[3]

  Users can set the length in one-second steps from 3 to 15 seconds.

• Confirmed sections[3]

  Google says it has dropped the price 30% from Gemini Omni 1.1, to $0.07 per second for 720p video.

• Confirmed sections[3]

  Users can draft at 360p or 720p, fix a detail with a follow-up edit instead of re-generating the entire clip, and then upscale the best results to 1080p or 4K.

• Confirmed sections[3]

  The model is generally available on the Gemini API and Google Cloud.

• Confirmed sections[3]

  Users can guide a video with character, product or style images, short motion clips of up to three seconds, and audio cues.

• Confirmed sections[3]

  Google says faces, garments, packaging typography and brand colors stay consistent across camera angles, and pitches the model for virtual try-on clips, catalog spins and branded video ads.

• Confirmed section title

  Google's Gemini Omni 1.2 Flash makes longer, cheaper AI videos

• Confirmed card[3] hook

  Google's new AI makes 15-second videos for 30% less.

• Confirmed card[3] point

  Gemini Omni 1.2 Flash costs $0.07 per second for 720p video.

• Confirmed card[3] point

  You can fix a clip with a follow-up edit instead of starting over.

• Confirmed card[3] point

  Google says faces, garments and brand colors stay consistent across angles.

• Confirmed card[3] take

  If you make video ads, AI drafts just got cheaper to try.

• Removed card[3] ask

  Would you post an AI-made video ad?

  name 'AI-made' not in cited source

• Confirmed why_it_matters[0]

  Two giants moved into everyday tasks this week: Claude now reads bank accounts, a step ChatGPT took in May, and Google cut the price of AI video by 30%.

• Confirmed why_it_matters[0]

  The race is moving from chat windows to tools people use with their own money and their own customers.

• Confirmed why_it_matters[1]

  Meta's warning shows that the AI race also runs through local politics.

• Confirmed why_it_matters[1]

  If voters push back on data centres, the build-out every new model depends on could slow, and companies will have to show towns what they get in return.

Checked but not used

These 1 passed the checks but were left out of the final post (for example a spare summary bullet), so they are not counted above.

• Not used what_happened[3]

  Google's Gemini Omni 1.2 Flash makes 15-second AI videos and cuts the price by 30%, to $0.07 per second at 720p.