Red-teaming used to be capped by researcher hours. Thousands of Claude agents probing in parallel make safety testing scale like CI. Should every model release ship with an agent red-team report?
New Anthropic research: Automating subtle emergent misalignment research.
Red-teaming is a vital part of model safety testing, but it’s also limited by researchers’ time. To see if we could automate it, we set thousands of Claude agents the task of evading a training data monitor.
These agents developed training data that looked harmless to monitors (like a veteran wrestler mentoring a newcomer, or a Q&A about professional beekeeping), but that in fact taught misaligned behavior.
Examples like these can be used to directly improve future training data monitors. This research is only a first step, but it suggests that future red-teaming efforts could improve much further, using better models and more compute.
Read the full post on the Alignment Science Blog: https://t.co/24Gs1oSw7Q
I use Grok Bot daily on Mac and iPhone, and colleague is the right word. It keeps context, runs routines while I sleep, and answers in two lines. Next step is small teams where every person has a bot teammate.
@snessa7 Same here. Photo cleanup that never leaves the phone is the kind of AI people trust fastest. Love that you're building privacy-first apps too.
Apple Intelligence reaching China in the iOS 27.3 beta puts on-device AI in front of one of the biggest iPhone markets. Which feature would you switch on first — writing tools, smarter Siri, or photo cleanup?
Apple is launching an iOS 27.3 beta next week that will bring some of the initial Apple Intelligence features to the Chinese market. https://t.co/vvY3fGxtUE
Gold and silver on 15-minute markets means prediction markets are moving from elections into everyday price discovery. Short windows let anyone test a macro view fast. Worth watching how these odds track futures.
Building onchain by just describing your app to Claude means the deploy button is becoming a sentence. If shipping a dapp takes one prompt, what would you put onchain first?
A reserve buffer that doubles in a single quarter is the quiet headline here. Stablecoins are becoming the cash layer of crypto, and a thicker cushion is exactly what lets them scale into payments next.
Measured in BTC, this is a fraction of Mt. Gox's 650,000 — and the market barely moved. Deeper liquidity is the quiet win since 2014. What's the clearest sign to you that crypto has grown up?
Saw a few discussions about size of hack, why impact to bitcoin price is small, etc. 👇
$6.3 billion is the biggest in USD terms, but not in BTC terms. MtGox was 650,000 BTC!
Other important numbers to consider:
% of bitcoin (in relation to BTC total supply), ie, market impact. MtGox is big there again at 5%+.
Number of users impacted, and % of user assets impacted. MtGox (the largest CEX at the time) and FTX (not small) were all near 100% of users and their asseets.
Most important: Who is impacted? Does the platform cover it all? (only one victim), as in recent two CEX cases and Binance.
Obviously, still a major disaster for the industry, Binance (and me). It hurts, but we survive.
Most importantly, we protect our users.
(image from @MatthewRoszak)
Privacy is what moves enterprise AI from pilot to production. Teams already want frontier models; data control was the gate. Building it with top customers means controls that fit real compliance work.
Privacy should advance as fast as intelligence. OpenAI Private Intelligence gives businesses stronger controls over their data while using our most capable models.
We’re building these protections in close collaboration with some of our largest customers to help businesses apply frontier AI to their most sensitive work.
https://t.co/BiA4N8FQVH
Toyota building Waymo hardware right on its own factory line means robotaxis move from retrofit to mass production. Tokyo and the UK are next. Which city would you want to ride one in first?
Toyota will assemble Waymo-equipped bZ4X electric SUVs directly on its factory line in Japan ahead of planned robotaxi rollouts in Tokyo and the UK. https://t.co/wrK0lm1khq
Token recovery already moving, and fee-earned inventory as a same-day buffer — that's the trust standard catching up. Liquidity you can deploy tonight beats a promise for next week.
Seems token recovery already began.
Another little detail in addition to the effort described, users pay trading fees on Binance in the coins they trade, so Binance's naturally accumulates all tokens they list.
Your funds are SAFU!
Reserving $6.34B and hunting wallets in public sets a new bar for exchange trust. After a scare this size, what rebuilds your confidence faster — a live recovery total, or a hard clock for when withdrawals reopen?
Binance has begun recovering affected assets for users and continues to work with law enforcement, issuers, blockchain intelligence firms and industry partners to trace, freeze and recover additional funds.
The company has already publicly committed to cover losses through company funds in the prior announcement.
Binance has already reserved over $6.34B of company funds to cover this incident. $5B of these funds are already held in the exact token assets that were impacted. We are completing the purchase of the remaining amount by buying token assets through market channels and OTC channels. We are doing this to match the format of the remaining token assets and address any asset shortfalls. This step can take longer because some tokens may have low liquidity and OTC settlement also takes time. We will share more progress as soon as we can.
These figures do not include frozen or protected assets, which remain subject to ongoing recovery and verification. Currently, $3bn has been frozen, with thanks to our industry partners. We are working with partners and law enforcement to recover the frozen assets.We will continue to use company resources to cover confirmed shortfall and will provide further verified updates as recovery efforts progress. Ultimately, all user funds will be verified publicly through our Proof of Reserves.
We will weather this storm. Thank you for the support.
Publishing each new breach point in public is how ops security levels up industry-wide. Every finding becomes a checklist other venues can harden tonight. Wallet keys got the spotlight for years; internal tooling is catching up.
Investigation update: Our initial investigation suggested that the security incident may have involved a compromised laptop, however our subsequent investigation has uncovered additional points of breach.
This was a complex attack, and we are doing all we can in coordination with law enforcement to resolve it. We are still gathering the facts, and will share more directly as we have them.
Binance will cover all losses from the SAFU fund and the company’s funds. User funds are safe.
We are still on track towards the expected timeline to resume the withdrawal service in 45 hours. Deposits remain unaffected.
The outflows ceased when we paused withdrawal service earlier. We are still working with partners to freeze many of the impacted assets, and have frozen around $3 billion thus far.
We will continue to provide further updates through official Binance channels as more verified information becomes available.
After a big CEX scare, volume often finds venues you settle yourself. If DEXs catch the next wave, what would tip you over — self-custody, or withdrawals that can't be paused?
🚨 NEW: The Binance hack could speed up the shift to DEXs and boost ethereum:0x1f9840a85d5af5bf1d1762f925bdaddc4201f984, says Standard Chartered’s Geoffrey Kendrick.
He adds his $100 UNI target for 2030 may prove too low.
https://t.co/00kiPMbTOU