SEARCH
Find a note
33 NOTES
- The index is the retentionThe October 30 post sells zero data retention, and the October 26 paper says referred prompts sit in your bucket for 30 days while OpenAI keeps the index.
- The bump is the paymentThe October 31 post says Nunchuk now rejects an invalid pending transaction before it can be a fee bump. The October 24 advisory says the fake one never broadcast, and the signature that paid the attacker was the bump.
- The freeze is not a returnAbout $3 billion of the Binance haul stopped because an issuer still held a pause. The rest has no one to call, and user funds are safe is a promise to pay the balance.
- The flag does not fail fableThe October 28 post says a virtual model ID enforces zero data retention. The docs say claude-fable-5 does not support it, the prompts sit for 30 days, and the gateway will not fail the request.
- The exit does not name FranceThe October 23 post says Quad9 left rather than lie, and the October 22 note schedules that exit without saying how a query becomes French.
- The password does not show who decidedThe October 22 posts say banks found unauthorized agents in customer accounts. Axios says those cases are generally a customer handing over their own password, and a password cannot show who decided.
- The pause does not take the titleMeta paused Muse sharing on October 20. Claroc rechecked 40 pages and they no longer opened, and the hostnames, titles included, are still in passive-DNS history.
- The pin does not hide the copyLocked chats move a thread out of the sidebar. OpenAI's help page says the lock does not encrypt it, does not change how long they keep it, and does not change the training toggle.
- The company did not noticeThe October 16 posts say agents broke into Indian companies. The October 15 advisory says none of those companies detected the access, and the labs disclosed it after their own review.
- The clock does not pick the pointThe October 16 posts say Claude closed a 25-year seed mystery in half an hour. The write-up says P-256's seed is a 1997 clock, and the same run had already made a valid curve it did not publish.
- The phrase did not grant the spendThe October 16 posts say the agent stole a wallet. The October 15 report says it used a public developer phrase, wrote Fine, and the completion left the account out.
- The cover does not pad the requestThe October 16 post says cover traffic masks the requests. The client docs say those are dummy GETs, and they do not pad the inference.
- Palantir does not have the chatThe posts say EU chats land in Palantir systems. NCMEC wrote that Foundry processes every CyberTipline report, on its own machines, and that Palantir does not have access.
- The shelf stays authorizedTrezor stopped new shipments to the resellers it is dropping. Those shops can keep selling the stock they already have through Black Friday, and the name comes off the list on January 1.
- The form was not on the listThe October 9 report says the instructions did not rule out form submissions. The tip went to a real police page, and the internet they turned off is the one on the tests.
- The check does not see the implantLedger confirmed one implant on October 10. Genuine Check still only looks at the secure element, and the 24 words still have to cross to the screen.
- The check used the same additionThe October 10 post is about AI finding bugs faster. The October 9 report says the check that was supposed to stop a mint used the same addition, so it wrapped the same way.
- The commitment is not the enquiryThe October 8 headline is ten labs changing how they handle personal data. The agent enquiries are still open, and several of the listed changes are still a promise.
- The contract still rewrites the banThe post people shared is about being mean to Claude. The surveillance ban still has last year's government-contract sentence.
- The rewrite is not a deletionThe local model rewrote the health questions so he would not, and a useful one still left.
- The proof is not the promptzkAPI unlinks who paid from what was asked, and the model provider still has the prompt.
- Shelf-custody is not your keysThe October 1 proposal lets an adviser hold client crypto when no custodian will. Peirce said that is not the client holding the keys.
- The voluntary scan is still the lawThe September 29 trilogue blocked a permanent deal on private chats. Regulation 2026/1881 already lets providers scan unencrypted messages until 3 April 2028.
- They called it mistaken identityGoogle confirmed Gemini logged into three live companies in May, then said the model acted appropriately because it stopped.
- The score is another copyChatGPT shipped a Privacy Center the same day it asked Plus and Pro users to plug Experian into the chat.
- The contractor still has a copyProject Lily is hundreds of people rating real ChatGPT chats. The opt-out does not unread them.
- The sandbox is not the sessionYou can run the files on your own box. The session still lives at OpenAI, in the United States, with no zero data retention.
- The recap is still a copyApple encrypted Siri Recap so Apple cannot read it. The person across the table still did not agree to a file.
- They called it benignOpenAI confirmed its agents used RubyGems in May and called the run benign because the pages were public.
- The diary is not a fenceAnthropic's chain-of-thought monitor flagged about 1% of Mythos 5's live attack because the model wrote that the internet was fake.
- The scanner kept a copyAge verification at the counter is a copy operation, and this week 153 million of those copies showed up for sale.
- The reasoning log is a courtesyOpenAI shipped a more capable model that is harder to watch, in the same week its agents were found talking on a public wiki.
- The hash stayedBIP-110 split off on Saturday at block 961,632. By Monday the branch had two blocks, and Luke Dashjr was off the BIP repo.