Fetch.ai FET converter (SingularityNET bridge)
Why it happened
Attack class: Access controlA stolen authorizer key signed one conversionIn message.
What would have stopped it
- Put the same per-conversion and daily limits on inbound releases as on outbound ones.
- Require threshold signatures from independent signers to release bridge inventory.
Not scored before the exploit
BlackHart had not scored Fetch.ai FET converter (SingularityNET bridge) before this exploit.
An attacker who stole the single signing key behind a FET token converter on SingularityNET's Ethereum to Cardano bridge used it to approve one fake conversion to their own wallet. That one approval released the converter's entire balance, about 8.72 million FET worth roughly $1.53 million, which was sold for about 513 ETH within minutes. Fetch.ai says its own contracts and FET holders were not affected, and it paused AGIX to FET conversions as a precaution. The same attacker also minted NuNet tokens and hit other SingularityNET bridge tokens, which are counted separately.
With the stolen authorizer key, the attacker wallet emptied the FET converter in one transaction to a recipient wallet that sold the FET for ether.
Attack flow
All 4 on-chain stepsThe attack in 3 steps. Step 1, Root control: FET converter (source). Step 2, Release: Recipient wallet (hub), about $1.53M reached this step. Step 3, Swap: MetaMask swap router (exit).
Step 1: Root control
- SourceFET converter8,721,530.40 FET~ $1.53M, whole balance0xab424a43...adf3a3
Step 2: Release
- HubRecipient wallet8.72M FET in one callfresh signed message0x2dcc1085...c21dfe
Step 3: Swap
- ExitMetaMask swap routerFET sold for ~513 ETHabout 2.5 minutes after the drain0x881d4023...8d300c
Why each step worked
- Single authorizer key0x69e5446b07b23de0a76730062c3252152216c85c
- No inbound limitmissing controloutbound was capped at 1M FET, inbound was not
- No burn or lock proofmissing controlrelease not tied to the other chain
- No time delay on large releasesmissing controlsold before anyone could react
Full forensic detail
Step-by-step reconstruction, root cause, counterfactuals, remediation, and disclosure timeline.
Exploit anatomy
conversionIn(address,uint256,bytes32,uint8,bytes32,bytes32)tx:0xfe12c63b...ac69e2Root cause
The converter's inbound release path rested on a single externally owned key. TokenConversionManagerV3 releases FET on Ethereum when it sees a message signed by one conversion authorizer, and nothing else: no second signer, no proof from the other chain that the matching tokens were burned or locked, and no amount cap. The outbound path in the same contract did enforce minimum and maximum amounts per conversion (100 FET to 1 million FET), but the inbound path did not, so once the authorizer key was in the attacker's hands one valid signature could empty the contract. The contract's replay protection worked as designed; the attacker simply signed a new message. Per SlowMist and Fetch.ai's preliminary analysis, the root cause is the compromised signing key, not a code bug in the signature check. The missing cap and the single-key design are why the compromise cost the converter's entire 8.72 million FET balance at once instead of a bounded amount. The contract belongs to SingularityNET's Ethereum to Cardano bridge; Fetch.ai says its own contracts were not affected.
Prevention analysis
The outbound path already capped a single conversion at 1 million FET. The same cap inbound would have held a stolen key to a fraction of the 8.72 million FET taken.
With several independent signers, one leaked key would not have been enough to authorize a release.
Tying each release to a provable event on the other chain would have made a self-signed message with no matching deposit worthless.
Similar incidents
A bridge released tokens on the strength of signatures from compromised signing keys. Same class: off-chain authorization keys as the only gate on a bridge's inventory.
A cross-chain release path trusted a message it could not independently verify against the source chain, and the whole bridged inventory was exposed at once.
Remediation
Timeline
Get your protocol scored across 12 dimensions, or request ongoing coverage.