AI Agents Are Starting to Pay for Things in Crypto
- Yahoo Finance framed the question bluntly: AI agents are starting to pay for things in crypto — and asked which rail wins, XRP, Solana, or USDC.
- Pantera's founder gave the thesis a one-liner: "They can't open bank accounts" — meaning agent-driven crypto demand is less ideology, more plumbing (Yellow.com).
- A report cited by Yellow.com counted 17,000 AI agents hitting DeFi since 2025 — but found humans still win at trading.
- Separately, SecurityWeek reported prompt injection attacks tricking AI agents into making crypto payments — the security model is, charitably, still forming.
The pitch is "agents need money rails." The reality is agents need money rails and a threat model. Right now they have one of two.
The Agent Stack Is Getting Audited — And Failing
- IBM Research and UC Berkeley ran 310 ITBench SRE traces across Gemini-3-Flash, Kimi-K2, and GPT-OSS-120B using the MAST failure taxonomy.
- Frontier models fail cleanly (
2.6 failure modes per trace); large open models suffer cascading failures (5.3 per trace), where one early reasoning mismatch poisons the context. - The strongest predictor of failure across all models: incorrect verification — agents "declare victory" without checking ground truth.
- Kimi-K2 showed a +46% spike in premature termination and +43% in "unaware of termination conditions."
- The recommendation is almost boringly sensible: externalize verification, require hard tool evidence before exit, and put termination control outside the model.
Every "autonomous agent" pitch deck should now be legally required to include its MAST failure-mode count. Most would not survive the slide.
Anthropic Hires a Referee, Trump Wants an AI Force
- Anthropic picked Accenture as its first embedded evaluator to help execute CEO Dario Amodei's AI-slowdown proposal — red-teaming models, alignment assessments, safeguard testing (CoinTelegraph).
- Both sides expect to invest at least $1 billion each over five years; Anthropic will fund it directly because no independent funding system exists yet.
- Amodei's stated worry: recursive self-improvement could "outrun our ability to understand and control these systems." Sam Altman and Elon Musk responded positively; Jensen Huang did not, arguing regulation isn't necessary.
- Meanwhile, President Trump said he plans an "AI Force" and an AI czar, posting that only "High I.Q. individuals need apply" — with no clarity on whether it's military, civilian, or a department (CoinTelegraph, citing Newsweek, NYT, BBC).
One company is paying a consultancy $1B to slow itself down while the government forms a vibe-based task force. The AI governance stack is being built in real time, badly.
Meanwhile, the Actual AI Research Keeps Shipping
- Hugging Face's MoE explainer is the useful one: gpt-oss-20b has 21B total parameters but only ~3.6B active per token (4 of 32 experts) — and ran at ~115 tok/s on an M3 Ultra, close to the back-of-envelope ~111 tok/s estimate.
- GGML and llama.cpp joined Hugging Face, with Georgi Gerganov's team keeping full autonomy; the stated goal is making local inference a "meaningful and competitive alternative to cloud inference."
- Transformers.js v4 shipped on NPM with a C++-rewritten WebGPU runtime, enabling accelerated models in Node, Bun, and Deno — plus a claimed ~4x speedup for BERT-style embedding models.
- Hugging Face launched Community Evals, letting benchmarks host leaderboards and the community submit reproducible results — an explicit shot at black-box leaderboards.
"Decentralized compute" is a crypto narrative. Local inference is the thing actually happening — and it's shipping on NPM, not on a token.
The Rest of the Tape
- The SEC's five-year innovation exemption for tokenized U.S. stocks has analysts naming Coinbase, Robinhood, and Circle as early winners (CoinDesk).
- Grayscale's Zcash ETF (ZCSH) filed for a 3-for-1 forward share split, effective after the Sept. 28 close, after ZEC ran roughly 2,800% over the past year (CoinTelegraph).
- Vitalik Buterin said AI will replace crypto wallets and interfaces (CoinMarketCap) — a headline that will be cited by at least 400 threads this week.
- And yes, there are now "Claude predicts LINK +300%" and "Gemini predicts XRP" stories. An LLM guessing at price targets is not analysis; it's a horoscope with a GPU bill.
"The AI tag is free," per one TradingView headline. It won't be for long — and most of the projects wearing it have no compute, no model, and no users.
Closing Take
The AI x crypto trade is consolidating around one real question: who pays the agent, and who verifies it? Agent payments are finding crypto rails because banks won't touch them — that's a genuine wedge. But prompt injection draining wallets and agents declaring victory without checking ground truth are the same failure mode wearing different clothes. Until verification lives outside the model, "autonomous agent" is just a liability with a token ticker.
Not financial advice. Side-Eye Signals is not liable when your agent pays the wrong address.