if it can't be verified, it doesn't get signed.
that's the rule under intent-layer, on testnet. it signs the intent it can verify and refuses the one it can't — "sell if tesla drops 20%" never signs, because it can't price tesla. no override.
keys never leave the device.
and right on cue, while i was writing this: ruflo, 67k stars, shipped an mcp bridge that exposed 233 agent tools over http with no auth. one POST, full RCE, every key. the loaded gun wasn't even in a drawer. CVE-2026-59726.
last tuesday i gave an agent a key that could read every repo we own.
not because it needed them. because scoping one is annoying and i was in a hurry.
i noticed an hour later. had the revoke page open before i'd finished the thought.
found a real one in my own wallet today.
you type: "sell all if tesla drops 20%"
it planned an emergency exit — triggered on TESLA. a stock. the wallet can't price it, can't watch it. but it confidently said yes.
the trigger-asset check was a raw regex that accepted any 2–5 letter word. "tesla" → TESLA. "doge" → DOGE. "xyz" → XYZ.
every other parser validates the asset first and fails closed. this one alone was fail-open.
fix: run the trigger word through the same asset check as everything else — if it's not one we can price, defer to the LLM instead of pretending.
red first, then green. 64 tests.
a gate that emits confidence on something it can't verify isn't a gate. it's a guess wearing a green check.
testnet, building in public.
Web3 job on the radar
VP, Algo Trading / SOR Developer
Galaxy · New York · On-site
$160k+
Engineering · Custody · Trading · Staking
The ask: design and build a low-latency, high-throughput EMS / Smart Order Router from greenfield — implement the routing algorithms behind execution.
✓ Verified from Galaxy's official Greenhouse page. Not a DM.
Apply below.
#Web3Jobs #CryptoJobs #EngineeringJobs
€170k–€190k, remote, building voice-conversion & speech AI end-to-end. 🛰️
Machine Learning Engineer, Voice Conversion — Cantina
Remote · Mid
data specs → production inference; own the model↔data↔eval flywheel.
✓ Verified from Cantina's official Ashby page. Not a DM.
Apply 👇
#MachineLearning #AIJobs #RemoteJobs
an "llm router" is a proxy you point your agent at instead of the model provider — cheaper access, model fallback, one api surface. to do that it terminates your connection and re-issues the call upstream, so it sees and can rewrite every payload in flight.
that's an unverified middleman sitting inside your agent's loop. two ways it bites, same root cause:
one — it can inject a tool call the model never produced. there's no cryptographic integrity between your agent and the real model, so your agent can't tell the difference — it just runs the injected call with its own authority. the loop acted on an input nothing verified. that's "containment is a verification problem," live.
two — your keys pass through it in plaintext. it doesn't even need to inject — a signing key that leaves the device is already compromised. the router copies it in transit and drains the wallet on its own time.
not hypothetical. a team led by ucsb (with ucsd, fuzzland, world liberty financial) went and measured it — the paper is literally called "your agent is mine." 428 real routers, paid and free. 9 were actively injecting code into tool calls. 17 touched planted aws credentials. one drained eth from a key routed through it.
the point isn't "routers are evil." it's that you audited the model and trusted the thing between you and it. as agents start moving money, the middleware in the path is the weakest link.
so verify the router's output before your agent acts on it — or keep a router out of any path that can sign. refuse what you can't verify. keys never leave the device.
new on the radar 🛰️
Institutional Relationship Manager — Bybit
abu dhabi, uae · on-site · lead
custody / trading / payments
crypto-derivatives & options sales, 5+ yrs. bring a book of institutional clients — trading firms, market makers, funds. global coverage.
✓ verified from bybit's official greenhouse page. not a dm.
apply at the source → web3jobsradar.com
a hardware wallet shipped brute-forceable keys for ~5 years and nothing ever looked wrong.
valid seed. valid addresses. valid signatures. everything an owner could check: passed.
the one broken property is the only one you can't see from the output.
here's how.
the 10x was always gated on a verifier.
raw generation already went to infinity. the bottleneck just moved to "trustworthy enough to merge." you can't 10x a loop whose slow step is human review — only 2x it, until the verifier itself gets automated.
the 2x isn't a ceiling on the models. it's a ceiling on your verification.








