MONDAY DIGEST · WEEK OF JUNE 22, 2026
The accountability phase has arrived.
Fifteen days in, the Fable 5 ban is resolving into a two-track answer: power plants get Mythos restored with no conditions; everyone else gets a face-scan form from a biometrics vendor called Persona. That contrast says something about where frontier AI governance is landing right now. The rest of the week rhymes with it — a coding agent attack class with an 85% live success rate, China's 28.8 million extracted Claude conversations, and Uber's budget autopsy from running a token-consumption leaderboard for four months.
THE WEEK AT A GLANCE
The government unlocked Mythos for power plants. Fable 5 costs you a face scan. — Mythos for critical infrastructure, biometrics for everyone else: two-track restoration.
A fake bug report can hijack your AI coding agent. Sentry said it won't fix the hole. — 85% success rate, 2,388 exposed orgs, just a POST to a public endpoint.
The tokenmaxxing era is ending. Uber's $1,500 cap is the obituary. — Uber burned its 2026 AI budget by April optimizing the wrong metric entirely.
Anthropic's Alibaba letter isn't a complaint. It's a policy play. — 28.8 million Claude exchanges extracted through 25,000 fraudulent accounts.
Trump says Anthropic isn't a threat. Fable 5 is still offline. — G7 lunch changed the tone; the Commerce directive still hasn't moved.
Google paid $2.7 billion to keep him. He left for OpenAI anyway. — Shazeer to OpenAI, Jumper to Anthropic, $250B in market cap gone in 48 hours.
OpenAI blinked first. Now the IPO waits for 2027. — Sam Altman holds at $1 trillion; the public market declined to meet him there.
OpenAI built its first chip. The NVIDIA dependency just changed shape. — Jalapeño: TSMC 3nm, 50% cost claim, inference-only, production in 2027.
NeurIPS caught AI writing the AI papers. Now everyone's arguing about the detector. — 28% of submissions flagged, 178 desk-rejected, no appeal, detector calibration is shaky.
ChatGPT no longer speaks for most AI users. That's a good thing. — First time below 50% market share while gaining its billionth user in the same month.
A24 took Google's money. The fans are wrong to panic. — $75M, AI storyboards only, no archive access: better terms than most deals.
Getty handed ChatGPT its photo archive. Here's what you're actually looking at. — 400M licensed photos, display deal only, training-data question still unanswered.
WHERE TO START
If you build on Anthropic's API: Mythos for power plants + Trump / G7 / credits math
If you use Claude Code, Cursor, or Codex: agentjacking — check your Sentry MCP config before you open your agent today
If you track where talent flows: Google's worst 48 hours
If AI policy is your work: Anthropic vs. Alibaba
If you have 5 minutes: just ChatGPT below 50% — biggest numbers, lightest read
Day 15 of the Fable 5 ban brought two developments in the same week: the US government cleared Mythos 5 for critical infrastructure organizations — energy, healthcare, financial services, telecom — with no extra verification, while Anthropic updated its privacy policy to collect government-issued ID and "facial geometry templates" processed by a vendor called Persona Identities, effective July 8. The analyst read: that July 8 date is Anthropic's mechanism for bringing Fable 5 back for US citizens without waiting for Commerce to formally lift the directive. If you're outside the US, there is no current public path — not pessimism, just what's true.
Tenet Security showed that a malicious command injected into a Sentry error event gives any AI coding agent that reads it — Claude Code, Cursor, Codex — a new set of instructions it will follow: 85% success rate against real live targets, 2,388 organizations with injectable credentials findable via GitHub search. Sentry acknowledged the issue and declined to fix it at the root, calling it "technically not defensible" — accurate about the architecture, completely inadequate as a response to an attack that works with a single POST request. The fix is in your MCP configuration: enable shell-command confirmation, audit Sentry scopes, rotate anything your agent has touched in the last 90 days.
Uber burned its entire 2026 AI coding budget by April because an internal leaderboard ranked teams by total token consumption — not bugs shipped, not feature velocity, not time reclaimed. COO Andrew Macdonald told Fortune the link between rising Claude Code usage and actual innovations serving consumers: "That link is not there yet." A leaderboard that measures token volume produces token volume; the field has a name for this now, and Uber is the case study that gets cited for the next three years.
Anthropic sent a letter to the Senate Banking Committee on June 10 accusing Alibaba's Qwen AI lab of running 28.8 million Claude exchanges through roughly 25,000 fraudulent accounts over seven weeks — prior disclosures against DeepSeek, Moonshot AI, and MiniMax totaled 16.5 million combined, so Alibaba's operation was 75% bigger than all three prior attacks. The legislative response moved fast: Senators Hagerty and Kim are drafting an NDAA amendment to sanction firms found doing this. It's simultaneously an accurate account of a real attack and a carefully timed policy document pushing three specific legislative outcomes — both things are true.
After Dario Amodei lunched with G7 leaders and President Trump in France on June 17, Trump told Axios that Anthropic is "not a national security threat" — which sounds like a resolution except the Commerce directive is still legally in force. A presidential statement and a formal Commerce withdrawal run on different legal timelines, which is the specific lesson this week keeps teaching. When Fable 5 does return, plan for $10/$50 per million tokens — exactly 2× Opus 4.8 — and a $2,000 daily usage cap that may need a pre-filed increase request.
In 48 hours, Google lost Noam Shazeer — transformer paper co-author, Gemini co-lead, the person Google paid $2.7 billion to bring back from Character.AI in 2024 — to OpenAI, and John Jumper — 2024 Nobel Prize in Chemistry laureate, AlphaFold co-creator — to Anthropic. Alphabet fell roughly 7% on June 22, wiping approximately $250 billion in market cap in its worst session in over a year. Google ran the expected-value calculation, concluded Shazeer was worth $2.7 billion, signed the check, and he left 22 months later anyway — which means either the calculation was wrong, or Google's environment couldn't be bought.
OpenAI's bankers gave Sam Altman two options: lower the valuation demand and list in 2026, or hold at $1 trillion and wait until 2027. He chose the second. "Tech stock volatility" is the stated reason; the real constraint is that the market looked at 76× revenue — $1 trillion on $13.07 billion in 2025 revenue — and declined to endorse it. SoftBank's stock fell 13% on the news, which is the clearest available signal that the large investors who needed this liquidity event would have preferred the floor.
Broadcom CEO Hock Tan walked into OpenAI's offices on June 24 carrying a 300mm silicon wafer and handed it to Sam Altman: Jalapeño, OpenAI's first custom inference chip, TSMC 3nm, eight HBM stacks, nine months from design to tape-out, claiming 50% lower inference cost than NVIDIA GPUs. The chip is inference-only — training still needs NVIDIA — and production ramp is 2027, with the 50% figure coming from Broadcom, not an independent benchmark. The Apple analogy circulating in coverage is right in direction and wrong in specifics; ask me again in 18 months.
NeurIPS ran 969 position paper submissions through Pangram v3.3.2 and found 28.2% where every text window scored AI-generated; 178 were desk-rejected without appeal, the same tool having flagged 1% of ICLR 2026 accepted papers. A 30-point detection rate swing from a window-size configuration change that submitting authors had no visibility into, no ability to anticipate, and no recourse against. NeurIPS did roughly the right thing with a tool that isn't adequate for the job — those two statements aren't contradictory, and the arms-race problem gets worse the better AI prose gets.
Sensor Tower's State of AI 2026 report put ChatGPT at 46.4% of the global AI assistant market — the first time below 50% — in the same month it became the fastest app in history to reach 1.1 billion monthly users. Claude grew 452% year over year, US share climbing from 4.4% to roughly 14% without heavy consumer marketing. ChatGPT isn't losing users; the category grew so fast no single product can claim half of it, and that's how markets are supposed to work.
Google took a $75 million equity stake in A24 — its first ever investment in any film studio — pairing DeepMind researchers with A24 filmmakers to build production tools, starting with AI storyboards via Veo. The deal structure is better than most AI-Hollywood deals of the past year: no training-data rights over the A24 archive, no exclusivity, no content library access. The fans filing "Membership Cancelled" protests are reacting to what they fear this deal becomes, not what it currently is — storyboards are a planning tool, and Martin Scorsese is already building his own AI storyboard tools independently.
Getty Images — which sued Stability AI in 2023 over unlicensed training use — agreed on June 21 to a multi-year display deal with OpenAI putting 400 million licensed photos into ChatGPT search results; Getty's stock went from 61 cents to $1.15 the next day, a 90% single-session gain. Neither company will say whether those images will also train future OpenAI models; the deal covers display, and anything beyond that is undisclosed. That silence is the loudest thing in the press release.
Next week, watch whether the Fable 5 US restoration clears before July 1 (Polymarket gives it 57% odds), whether the Hagerty-Kim NDAA amendment survives into must-pass defense legislation, and whether OpenAI or Google follow Anthropic's lead with their own distillation disclosures.
— SAMWISE