> daily_signal(2026_07_08)
Meta's new Muse AI can put your face in someone else's AI photo — on by default, no notification, ever — while a Russian phishing crew we've tracked for a week gets a $10 million bounty.
PickBits Daily Signal · Wednesday, July 8, 2026
// tl;dr
- Meta's new Muse Image generator opts every public Instagram account into AI face-remixing by default, with no consent and no notification, ever. Tag a public account in a prompt and Muse can use that person's photos to generate a new image with their likeness — no request, no consent, and Meta's own help page says no notification, ever. The opt-out is buried three menus deep and isn't retroactive.
- The State Department put up to $10 million on two Russian hacking crews, UNC5792 and UNC4221, for a phishing campaign against Signal and WhatsApp accounts. We flagged this campaign when Ukraine's security service went public on July 2; the reward is now official, and the FBI says stolen backup keys stay valid even after a victim starts a new account.
- Apollo's chief economist, Torsten Slok, published a chart showing zero AI profit-margin lift at the 493 S&P 500 companies outside the "Magnificent Seven." His read: regulated industries need years, not months, to capture AI productivity gains, and if that timeline runs to 2029, a lot of AI-premised valuations face a correction.
- Meta's Brain2Qwerty v2 reconstructs full sentences from brain activity recorded outside the skull — no surgery required. Word error rate dropped to 39%, the best participant hit 22%, and the non-invasive approach is closing the gap with surgical brain implants like the one that gave ALS patient Casey Harrell his voice back.
Meta had a rough month before today even started. We watched a dormant facial-recognition tool surface on fifty million phones, then an AI chatbot bug that let attackers hijack real Instagram accounts, and now Muse Image ships with your public photos fair game for anyone's prompt — no consent required, unless you find the toggle yourself. That's the privacy story today. The threat side moved too: the Russian social-engineering campaign against Signal and WhatsApp that we first flagged when Ukraine's security service blew the whistle on July 2 now has a ten-million-dollar bounty attached, official confirmation of a threat that's been live since March. And underneath both, Apollo's chief economist put a number on something a lot of people have been feeling but not measuring: outside seven companies, AI hasn't moved profit margins yet, and Wall Street is pricing in years of gains that may not land until 2029. Three different stories, one shared question none of them answer cleanly: who's actually accountable when the thing moving fast stays invisible until it isn't.
A default setting, a bounty, and a chart all landed the same message this week: check what you assumed was already handled.
1. Meta's Muse Image opts every public Instagram account into AI face-remixing, with no consent and no notification, ever.
A photo you posted for your friends just became material for a stranger's prompt.
Meta launched Muse Image on July 7. It's the first in-house AI image generator from Meta Superintelligence Labs, built to compete with OpenAI's GPT Images 2.0 and Google's Nano Banana 2. It's live now inside the Meta AI app, Instagram Stories, and WhatsApp, with deep integration into Instagram specifically. Tag any public account in a prompt, and Muse can pull that person's own photos to generate a new image using their likeness. No request goes out. No consent is asked. Meta's own help center confirms it: the person whose face was used will not be notified, under any circumstance.
The setting exists to turn off, but Meta buried it three menus deep: open Instagram, tap the three lines top right, scroll to Sharing and reuse, and toggle off both Posts and Reels. Even then, the fix only applies going forward. Anything already generated using your face stays out there, permanently. This is Meta's third privacy misstep this month alone — after a dormant facial-recognition tool we found sitting quietly on fifty million phones, and an AI customer-support chatbot that let attackers hijack high-profile Instagram accounts just by asking it to change the account email. Ship first, let users discover the cost later, is starting to look less like an accident and more like a habit.
techcrunch.com: Meta rolls out Muse, a new AI image generator (July 7, 2026)
digitaltrends.com: Meta's new AI can generate images of you from your Instagram, and you're opted in (July 2026)
2. The US just put a $10 million bounty on the Russian hackers targeting Signal and WhatsApp backup keys — a campaign we've been tracking since July 2.
A text that looks like tech support is how a spy agency gets into your phone now.
The State Department's Rewards for Justice program is offering up to $10 million for information on two Russia-linked hacking groups, tracked as UNC5792 and UNC4221, tied respectively to the FSB and Russia's Border Guard Service and military intelligence. We flagged this campaign on July 2, when Ukraine's security service went public with a joint finding alongside the FBI; by July 3, the reward was already in motion. Today it's official, and the FBI filled in how the campaign actually works. It's social engineering, not a software exploit: fake "official support" texts, doctored Signal group-invite pages, and prompts asking a target to hand over a verification code, a PIN, or a backup recovery key.
The backup key is the real prize. Once an attacker has it, it stays valid even if the victim creates a brand-new account with the same phone number, meaning a single successful phish can grant access that survives someone's attempt to start fresh. The campaign has been running since at least March, targeting current and former US government officials, military personnel, political figures, and journalists — and it's not the first time we've seen Russian state hackers run exactly this profile of target. A 2017 case against two FSB officers, indicted for the Yahoo breach, named the identical combination: government officials and journalists, both. Different breach, years apart, same playbook.
3. Apollo's chief economist ran the numbers on AI's profit lift outside the "Magnificent Seven." It's zero.
The market bought a story before it checked whether the numbers agreed.
Torsten Slok, chief economist at Apollo Global Management, published a chart this week making a point Wall Street has mostly avoided saying out loud: at the roughly 493 S&P 500 companies outside the "Magnificent Seven" tech giants, there is no measurable AI-driven margin expansion yet. His argument is about timing, not whether AI works: regulated, process-heavy sectors, like healthcare, banking, energy, pharmaceuticals, and manufacturing, need years of compliance work and workflow redesign before productivity gains actually show up on a balance sheet, not months. Even where the gains are real, Slok notes, they're often too diffuse to be captured cleanly in standard financial reporting, which delays the point where investors can verify anything at all.
The stakes are in the timeline math. Slok's framing: if the productivity bump takes five years instead of five months, a lot of AI-premised stock valuations are due for a hard correction, with market expectations and actual earnings potentially diverging out through 2029. This lands against a backdrop we've been tracking all year: Anthropic's $965 billion valuation and confidential S-1 filing, Alphabet's gen-AI revenue up 800% year over year matching its capex line, hyperscalers stacking tens of billions into AI infrastructure. The market has, in fact, been pricing in a lot. Slok's chart is the first widely circulated attempt to check whether the other 493 companies are actually cashing that check yet.
4. Meta's brain-reading AI just closed most of the gap with a surgical implant — no scalpel required.
The distance between a research lab and a working voice just got a lot shorter.
Meta's FAIR research team published Brain2Qwerty v2, a model that reconstructs full sentences from brain activity recorded entirely outside the skull, using magnetoencephalography to pick up the magnetic fields the brain's motor cortex gives off during intended typing. No implant, no surgery. Nine volunteers wore the sensor for ten hours each, silently mouthing out sentences on a keyboard they couldn't see, and the model learned to reconstruct what they meant to type from the signal alone. The average word error rate dropped to 39%, down from 55% for the raw signal with no language model helping out; the best individual participant hit 22%. A fine-tuned language model does the heavy lifting at the sentence level, turning noisy, error-prone character guesses into something coherent.
Until now, reliable brain-to-text meant an actual implant. We watched Casey Harrell, who lost his ability to speak to ALS, get his voice back through exactly that kind of surgical device back in June. It used electrode arrays placed directly into his speech motor cortex, reading a few hundred neurons at a time. China's NEO implant, the first invasive brain-computer interface cleared for commercial sale anywhere, beat Neuralink to market the same month. Meta's version skips the operating room entirely. It's still a research result, not a product. Real-time use hasn't been demonstrated yet, and a 39% word error rate is still rough for daily use, but the gap between a scalpel and a magnetometer just got meaningfully smaller, for a technology built for people who've lost the ability to speak or move on their own.
» What to watch this week
- Whether Meta updates Muse Image's default setting under pressure, or leaves it opt-out. This is the third privacy misstep from the company in a month; the pattern itself, not any single feature, is becoming the story worth tracking.
- Whether UNC5792 or UNC4221 members are ever identified, and whether the bounty produces a tip. A $10 million reward is a strong incentive against a state-linked group with FSB cover — worth watching whether that changes anything operationally, or just raises the campaign's visibility.
- Whether Slok's warning shows up in the next round of Big Tech earnings calls. If AI-adjacent companies start getting asked directly about margin timelines outside their core seven, that's the signal the market is starting to price this in rather than around it.
- Whether Brain2Qwerty crosses into real-time use. Meta's own team names the remaining barrier as timing alignment, not raw accuracy — closing that gap is the difference between a research paper and something a patient could actually use.
Tomorrow's signal lands here.