Research · 13 October 2026 · 6 min read
Bounding what an agent does is not bounding what it believes
By James Warder
- Oracles
- Price data
- Agent safety
A correct decision on a wrong number
On 22 February 2026 someone made one abnormal trade in the USTRY/USDC market on Stellar's built-in exchange. The market had less than one dollar of hourly trading volume and no other trades for more than ten minutes. The oracle used by a community lending pool computed a volume-weighted average price from that market. Because nothing else traded, one trade was the average. USTRY went from about $1 to about $106.
The attacker then posted USTRY as collateral and borrowed against it: first 1,000,196.70 USDC against 13,003 USTRY, then 61,249,278.31 XLM against another 140,000 USTRY. Losses passed $10M. The oracle's infrastructure was not exploited. It reported, faithfully, the only market there was. (QuillAudits)
Now put an agent in that pool's place: a yield agent that rebalances into whatever looks best, with a perfect spending limit. It would stay inside every limit and still walk the owner's money into a position priced on fiction. The limit bounds what the agent does. It says nothing about what the agent believes.
Three things to bound, not one
Looking back across the systems we have built, we kept finding the same gap in three forms:
| Bound | The question | What fails without it |
|---|---|---|
| What the agent may do | Which actions, how much, to whom, until when | The agent sends too much or to the wrong place |
| What the agent believes | Are the prices and states it reads true and fresh? | The agent acts correctly on a wrong number |
| What the agent walks into | What state is the venue in when the action lands? | The agent enters a pool mid-exploit or a market with no depth |
A mandate covers the first row. The other two need their own rules, and they are the ones that get forgotten, because a limit feels like safety.
The rules we hold ourselves to
Every one of these comes from an incident, not from caution.
1. Check the price before anything else. Every number in a trading decision is a priced number: expected value, slippage, notional size. Computing them before the price is validated means doing arithmetic on inputs you are about to throw away.
2. At least three independent sources, or refuse. With two sources, a median is just an average and tells you nothing when they disagree. Below three, the action is refused. We never fall back to a single feed, because the single feed is exactly what the attacker above used.
3. A bounded-move rule, separate from source agreement. Sources agreeing does not make a price right. They can all read the same thin market. A rule that caps how far a price may move between two observations catches a 100× jump that agreement never will.
4. Staleness per feed, measured from the publish time. Every feed has its own heartbeat. A single global threshold set below a feed's heartbeat halts constantly, until operators learn to ignore it, which is worse than no guard. And staleness must be measured from when the source published the price, not from when we last fetched it. A frozen upstream looks perfectly fresh to a service that only knows when it last asked.
5. Know what a halt stops. When sources disagree beyond a threshold, halt valuation and every action that increases risk. Never halt withdrawals, and never halt fee accounting. A halt that freezes fees while withdrawals stay open publicly announces an escape route. A halt that freezes withdrawals traps owners at the worst moment.
6. Confirm the feeds exist before shipping on a chain. "Never single-source" is a requirement about the chain, not just the code. If three independent feeds do not exist for an asset on a chain, the rule cannot run there yet.
7. Advice is not authority. Market data and model research can explain a decision, pause an agent or ask the owner to review. They can never raise a ceiling, remove a rule or override a contract's refusal. A provider outage should make context unavailable. It should never disable enforcement, revocation or withdrawal.
The correction in our own notes
When we first wrote these rules, our internal notes described this incident as a case where "every feed agreed" on the wrong price. Rereading the public write-up, that is not what happened. The oracle took its price from a single market. There was no disagreement to detect because there was only one voice.
The rule we wrote, a bounded-move check separate from source agreement, is still right. But the reason we gave for it was wrong. That matters, because a rule justified by the wrong incident gets argued away by the first person who reads the right one. We fixed the note.
One story we did not repeat
While researching this post we found a widely shared article about a "$45M AI agent exploit" in April 2026, in which agents supposedly traded correctly into poisoned price feeds. It would have been the perfect opening. It names no protocol and no agent, links no transaction, cites no post-mortem, and promotes the author's own service. We could not verify it, so it is not in this post. If you have a primary source for it, send it to us.
Where we stand
Gol's live profile today enforces one action, USDC transfers to listed recipients, which needs no price at all. The rules above are written into our specification for the trading and yield actions that come next, and they ship with those actions. We are not claiming them as live.
Disclosure: Morca Labs builds Gol. The incident in this post did not involve Gol.
Sources
- QuillAudits, YieldBlox $10M hack explained
- OWASP Smart Contract Top 10 (2026), SC03 Price Oracle Manipulation