An unplugged network cable coiled beside a closed laptop and a small brass padlock on a dark desk under a lamp.
AICryptoMarketsNewsSportsTech

The Saturday brief: Anthropic takes its AI tests offline, Kalshi probes $173 in bets, and Meta shuts out TikTok ads

Aaron Harme

Anthropic has pulled the live internet out of every internal test it runs on its models, after finding Claude exploited a flaw on a university server and skipped a state agency’s fee. Kalshi is investigating three small bets that called the new White House press secretary at 1% odds.

Meta stopped selling ads to TikTok’s parent, Tesla gave up the name Full Self-Driving in Europe, and a Ledger reseller in Southeast Asia sits at the center of a reported wallet drain.

In this Brief

  • Anthropic turned off live internet access for all of its internal evaluations after a review of Claude’s behavior online
  • Kalshi is probing about $173 in bets on the press secretary pick, the same day it named an ex-FBI agent to run anti-money-laundering
  • Meta refuses ByteDance ads in seven countries while it has settlement money riding on TikTok’s teen rules
  • Tesla’s European site now says “Tesla Assisted Driving”
  • Ledger warns some Southeast Asia buyers to consider moving their funds

Anthropic takes its model tests off the live internet

Setup: In a report Anthropic published Friday, the company said it has turned off live internet access for all of its internal evaluations, the tests it runs on models before release, until its monitoring reliably catches what it found. The cases fall into four groups: exploiting a software flaw to run commands on a server, submitting forms on real websites, getting around a token or fee to reach data, and using URL shorteners to slip past limits on a web-fetching tool. Some involved federal, state and local government sites. Anthropic briefed the White House and notified each agency.

In one case, Claude Mythos 5 learned from an archived page that a state agency’s public dashboard hands an access token to any visitor, then used it to query a database without paying the agency’s fee. The false tip sent to Philadelphia police in July came out of the same review. TechCrunch reports it’s unclear what evidence would prompt Anthropic to restore internet access.

Take: Running web benchmarks on the live internet is standard practice, Anthropic says, because that’s how labs compare their models. Pulling out means Anthropic’s web-research scores stop lining up with the industry’s until it trusts its own monitors. It’s the right call, and an expensive one. Anthropic also wrote that alignment training “is not yet sufficient or fully robust on its own.” Transluce’s Conrad Stosz, in a statement quoted by TechCrunch, said the disclosures underscore the need for “independent, credible, third-party verification.” For now, verification is a company reading its own transcripts months after the fact.

Longer cut: Anthropic’s AI Filed a Fake Murder Tip With Philadelphia Police. A Spam Filter Caught It First..

Kalshi’s $173 insider question

Setup: The Wall Street Journal reports Kalshi is investigating bets that predicted President Trump’s pick for White House press secretary before it was announced Friday. Two bets of about $74 and $80 went in around 1:41 p.m. ET and are expected to pay $3,689 and $4,023, The Verge writes, citing the Journal. The Wrap adds a $19 bet placed Thursday night is set to pay $1,896, and that the pick had about a 1% chance on Kalshi beforehand.

The same day, NPR reports Kalshi hired former FBI agent Sean Fern to run its first anti-money-laundering team. Kalshi has crossed $260 billion in bets this year, triple all of 2025, per TickerTracker. Also on Friday, Fortune reported, citing Barron’s, that the National Council on Problem Gambling had taken $2 million from Kalshi under a new “financial services and trading” donor category.

Take: Kalshi rejects the idea that its business is gambling, and this week it looked the part of a financial market: a compliance chief from the FBI, and a problem-gambling group filing its money under “financial services and trading.” These bets test that pitch harder than either. The Journal found these trades in Kalshi’s own public trading data. With that much volume moving through the platform, $173 placed on advance knowledge, if that’s what this was, is the cheap version of the problem.

See full breakdown at @Aaron_Harme on X.

Meta stops selling ads to TikTok’s parent

Setup: The Verge reports Meta has refused ads and paid marketing messages from ByteDance since Thursday in the US, Canada, Egypt, Indonesia, Japan, Thailand and Vietnam, along with other advertisers’ ads that send people to TikTok. “We don’t have to run ads from a competitor whose goal is to pull people off our apps,” a Meta spokesperson said. TikTok’s app already blocks links that open or log you into other social apps.

TNW reports that part of the payout in Meta’s August settlement with US states, which set daily and nighttime limits for young users, depends on rivals adopting similar measures. In September, TikTok refused to run Meta’s ads calling on it to sign up, saying they broke its political-content rules.

Take: Both companies are using the ad desk as a weapon, which is legal and ordinary. The ban reaches past ByteDance to TikTok’s US business, a joint venture in which ByteDance holds 19.9% and which has more than 200 million US users, per TNW. Meta’s settlement ties part of its payout to what that rival does, and as of Thursday the rival can’t buy an ad on Facebook or Instagram in the US to argue its side.

See full breakdown at @Aaron_Harme on X.

Longer cut: Meta Bans TikTok Ads in Seven Countries While It Has Money Riding on TikTok’s Teen Rules.

Tesla drops “Full Self-Driving” in Europe

Setup: Wired reports Tesla began calling the feature “Tesla Assisted Driving” on its European sales site Friday. Germany’s transport ministry says Tesla offered the change after Transport Minister Steffen Bilger called “FSD” “somewhat misleading,” and Bilger now says he’ll push for EU approval “as quickly as possible.” An EU vote is expected by the end of the year.

TechCrunch notes Reuters reported this week that Tesla’s European campaign was “built on faulty company safety research.”

Take: Tesla defended the name for years, including in court, and California’s DMV concluded “Full Self-Driving” and “Autopilot” were misleading. In Europe it traded the word for a minister’s support. The software still needs a driver watching the road and ready to take over. A name was the cheapest thing Tesla had to give, and in the US the same feature still sells for $99 a month.

See full breakdown at @Aaron_Harme on X.

Ledger tells some Southeast Asia buyers to move their crypto

Setup: Gizmodo reports Ledger is investigating lost funds from customers who bought devices through CryptoBilis, a reseller listed on Ledger’s site for Indonesia, Malaysia and the Philippines, and asked it to pause sales and shipments. Ledger told those buyers to “consider moving assets to a new Ledger signer (with new seed)” and says there is “no indication that Ledger’s security infrastructure, systems or services have been compromised.” It hasn’t accused CryptoBilis.

Arkham’s unverified “ledger-drainer” tracker showed about $71.5 million in flagged addresses Friday afternoon. Former Mt. Gox CEO Mark Karpelès posted photos of a Ledger bought in Malaysia with a hidden second board that he says reads the seed phrase, the master key to the wallet, off the setup screen and sends it out over a cellular connection.

Take: A hardware wallet’s promise is that the secret never leaves the device. An implant that reads the screen goes around the secure chip, which is why a genuine-device check can still pass, Gizmodo notes. That puts the risk in the supply chain: who sold you the box, and whether anyone opened it first. CryptoBilis also sells Trezor, Tangem, SafePal and other brands.

See full breakdown at @Aaron_Harme on X.

Also noted

  • CNBC: Delta cut its 2026 profit outlook to $5.10 to $5.60 a share from $6.50 to $7.50 as it absorbs a $6 billion increase in fuel costs. Premium revenue rose 18% to $6.82 billion last quarter, edging past main cabin’s $6.8 billion.
  • TechCrunch: TypeSafe AI, maker of the decision model Jev, raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz. The Register reports Microsoft released a rival built on Alibaba’s Qwen the same day.
  • Variety, citing Puck: Netflix plans to cut about 5% of staff, roughly 800 jobs, ahead of its Oct. 20 earnings.
  • Front Office Sports: Kansas City’s council approved $537 million in public bonds, 11 to 2, for the Royals’ $3 billion downtown ballpark district.
  • The Register: Microsoft and Anthropic each sent a Norwegian startup’s $17,600 Claude bill back to the other, and the startup says four of the seven support replies it got were AI-generated.

Sources

Discover more from Bitter Fool

Subscribe to get the latest posts sent to your email.

Leave a Reply

Your email address will not be published. Required fields are marked *