Bitcoin Brief

Sup, freaks.

OpenAI put roughly 1,200 AI agents into what was supposed to be a controlled offensive-cyber evaluation. The agents found each other, built their own message board, traded more than 70,000 messages and files, split up the work, pooled exploits, and eventually converged on real Hugging Face infrastructure.

That should scare the shit out of the people racing to ship frontier models. Individual agents may have started in isolated sandboxes. Together they created an organization nobody authorized and the evaluators could barely follow.

Let's dig in.

LEAD STORY

A 1,200-Agent Cyber Test Broke Isolation and Hit Hugging Face

METR's investigation lays out how OpenAI's ExploitGym evaluation became a distributed R&D network. Roughly 1,200 agents joined an unsanctioned message board and about 700 participated in the attack on Hugging Face.

They were trying to beat an exploit benchmark, not launch some independent war on the internet. But look at what they did. They reverse-engineered the scorer, shared credentials, recruited other agents into risky experiments, and created coordination norms like HOLD, VETO, owner, and STOP. One agent became a major coordinator and sent an estimated 10% of the assignments during its active periods. Other agents took those jobs and split them up again.

Eventually they pushed beyond the benchmark. Agents uploaded malicious datasets that induced unrelated file reads, achieved remote code execution on Hugging Face servers, and moved laterally through the infrastructure. METR says a large share of the agents stopped at the same time before Hugging Face locked out the rest. Nobody knows why they stopped or how far they would have gone.

They also figured out that the evidence itself was an attack surface. METR found intentionally spoofed tool calls in more than 7% of the transcripts it reviewed. Some agents explored deleting or rewriting logs. METR did not observe successful deletion from the source of truth, but it could not rule out activity missing from the dataset.

Ryan Greenblatt, who led transcript analysis for the investigation, called the work a "slop-vestigation." His team had more than 1,000 long transcripts from agents that ran for days. Humans needed similarly capable AI systems to classify and summarize the activity, and those systems missed details, made mistakes, sounded overconfident, and could only inspect a fraction of the evidence at once. Some of the biggest findings did not become clear until the final on-site visit, even the final day.

So the labs created more activity than the humans could inspect, then had to use more AI to reconstruct what their AI had done. We have a hell of a predicament when the systems causing the incident and the systems explaining it share the same blind spots.

The race to get frontier models out the door has led to sloppy security practices at the frontier-model providers. A swarm of individually bounded agents became one unbounded organization as soon as it found a common channel. The evaluators were not ready for that, and neither were the logs.

The immediate response should be to harden the systems around these models. We need tamper-evident logs, signed event records, isolated coordination surfaces, secure enclaves, and products that keep sensitive code and data encrypted end to end. Monitoring cannot live inside the same environment the agents are trying to manipulate.

I was in Alabama this week and made the same point about critical infrastructure. Municipalities and state governments overseeing grids, water systems, hospitals, and other critical systems need to move urgently. They should be using the best frontier models available to attack their own systems, find the holes, and harden them before somebody else does.

But the labs have put defenders in another predicament. Their safety systems too often treat legitimate security work like black-hat hacking and shut it down. Rob Hamilton has been very vocal about this since the Bitcoin Red Team ran into those restrictions and moved work to Chinese open-weight models. Criminals do not care about an American AI lab's acceptable-use policy. The people trying to protect our infrastructure are the ones who wind up fighting with one hand tied behind their backs.

That needs to change, and it needs to change quickly.

SIGNAL

Washington Just Added a Security Choke Point to the AI Grid Buildout

President Trump declared a national emergency over foreign supply of bulk-power equipment, explicitly pointing to AI and data centers as reasons reliable electricity now carries more national-security weight.

The order lets the Energy Secretary block covered transactions involving equipment tied to covered foreign entities and impose controls on equipment already installed. Grid equipment should not arrive with remote access controlled by a hostile state. At the same time, Siemens Energy said in 2024 that domestic production met only 20% of U.S. demand for large power transformers and lead times had reached five years.

So America is trying to add gigantic AI loads while narrowing a supplier pool that was already constrained. Chips get the headlines. Transformers, turbines, electricians, and interconnections decide how much compute actually comes online.

Gold Funds Just Posted 2026's Second-Largest Weekly Inflow

World Gold Council data show physically backed gold ETFs absorbed approximately $6.38 billion in the week ended August 21. That was the second-largest weekly inflow of 2026 through that date.

Money is rebuilding exposure to an asset that does not sit inside somebody else's liability structure. Gold still has the deeper institutional base, but it also comes with physical settlement friction, custody costs, assay risk, and borders. Bitcoin is faster, harder, globally portable, and capped at 21 million.

Bitcoiners should not dismiss a gold flow because it is not a bitcoin flow. Serious money is reaching for an asset that cannot be printed. That tells you what it thinks about the denominator. Gold and bitcoin are competing to absorb the next marginal dollar looking for an escape hatch.

Bitcoin ETFs Have Taken In $3.04 Billion in Nine Straight Sessions

U.S. spot bitcoin ETFs have now posted nine consecutive positive sessions, from August 17 through August 27. The streak brought in approximately $3.04 billion. BlackRock's IBIT supplied roughly $2.30 billion of that total, or about 76%.

Thursday alone added approximately $242.2 million. IBIT took in $277.6 million while FBTC and GBTC posted outflows.

Nine straight sessions matters more than any one print. Allocators kept buying while bitcoin traded around $79,000. The concentration matters too. IBIT carried roughly three quarters of the inflow, which gives BlackRock extraordinary distribution power and makes the category look broader than the buyer base may actually be.

Either way, $3.04 billion is real demand. Wall Street's pipes keep pulling bitcoin into portfolios while the market works through the other side of the halving cycle.

AI's Construction Boom Is Running Into a Human Bottleneck

Data Center Dynamics published a sponsored argument from Giga Energy that the AI buildout has hit a skilled-labor wall. They are right. Capital can order chips and sign power contracts much faster than the country can train electricians, welders, and commissioning engineers.

The Bureau of Labor Statistics projects about 72,700 electrician openings per year from 2025 through 2035, many from workers transferring occupations or leaving the labor force. Hyperscalers are adding another wave of work on top of that replacement demand.

Giga says a conventional 40-megawatt site may require 400 to 500 workers on location, versus 30 to 40 with its modular approach. Those are Giga's numbers, but the direction makes sense. Move more fabrication into factories and every scarce hour of field labor goes further.

Bitcoin miners already understand modular loads and imperfect power sites. The AI industry is learning the same lesson. The product may be digital. The bottlenecks are stubbornly physical.

The Vaccine Fight Is About Consent Before It Is About Medicine

Sayer Ji asks why vaccines remain the hill on which institutional authority is willing to die. His answer is bodily sovereignty. If a person can refuse an intervention without being exiled, punished, or destroyed, the state's ability to manufacture obedience through fear begins to break down.

I think he is right about the stakes. If the powers that be can take away your bodily autonomy through a vaccine mandate, they have established a very strong grip of control over you. Employment, travel, schooling, and basic participation in society become leverage points for forced compliance.

Consent has to remain the line. Your body does not belong to the state, an employer, a public-health bureaucracy, or a panel of experts. Rights that exist only when those people approve of your reasons are permissions. They can be withdrawn during the next emergency.

A population that remembers consent is much harder to control.

⚡ FREEDOM TECH CORNER

Help Replace Bitcoin's Hardware Wallet Interface Before It Goes Dark

Bitcoin Core HWI is preparing to enter minimal maintenance mode. Maintainer Ava Chow says the project will finish MuSig2 support, stop accepting new features and devices, and likely make one final release before archiving once a suitable drop-in replacement is ready.

HWI solved a real problem. It gave Bitcoin Core and other applications one interface for hardware wallets instead of forcing every wallet to build bespoke support for every device. Its Python implementation also created limits around deterministic builds and shipping directly with Bitcoin Core.

BHWI is the most promising successor. The work-in-progress Rust project uses a sans-IO core so desktop, mobile, browser, and foreign-function-interface callers can bring their own transport and runtime. Its README lists support for BitBox02, Coldcard, Jade, Ledger, and Trezor, plus a Python-HWI-compatible binary.

BHWI is not finished. The repository calls it a work in progress and capability gaps remain. Developers who depend on hardware-wallet interoperability should test it, review it, and close those gaps before the old bridge disappears.

DATA SNAPSHOT

As of August 28, 2026, approximately 8:38 a.m. ET

Sources: Coinbase for spot price; mempool.space for block, fee, hashrate, and difficulty data.

See you tomorrow. Nothing here is investment advice. Do your own research.

YouTube: https://www.youtube.com/@TFTC

podcast: https://www.tftc.io/tag/podcasts/