DSS 2026 / Day 2 / Amber Hall
Real-Time Threat Detection for DeFi with Stream Engines
Jason W
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:01Hello, I am Jason, an independent
0:04security researcher and deeply
0:06passionate about Web three security and
0:09build build the tools and a system to
0:12contribute back to the community.
0:15Uh today I will talk about real-time
0:18threat detection for D5.
0:21I'm focusing on two core goals. First,
0:24uh tag threat as early as possible.
0:27Um
0:29blockchain
0:30are almost always discovered after the
0:33fact. Funds are already drained that
0:37makes this a unique
0:39hard problem compared to the Web two
0:42security.
0:43The core question I am trying to answer
0:46is straightforward. Can we spot attacks
0:49fast enough to create a window for
0:52anti-money laundering
0:54and ideally even interrupt the hike.
0:58Second, build this real-time detection
1:01system under tight resources constraint.
1:05I run a R E T H execution node in
1:09minimal mode to keep overhead no.
1:12Let's walk through the system
1:14architecture and the data pipeline.
1:17I build a custom extension that collects
1:20transaction traces and the event
1:23logs. All raw
1:25elementary flows into a adapter
1:28component.
1:29This adapter normalized on-chain events
1:32and enrich it will enrich
1:35smart contracts, core graphs and
1:38attached and attached security labels.
1:41In that inside this adapter a 10
1:44propagation algorithm track
1:46mixer protocols such as Tornado Cash.
1:50It marks address founded through Tornado
1:53Cash and propagate that 10 label
1:55downstream.
1:58Normalized enriched the events then into
2:01an event pool.
2:03A stream matching engine consume this
2:05upstream event. I predefine a set of
2:08matching patterns and evaluate them
2:12against the live event stream. When a
2:16pattern matches
2:18the event stream will fire an alert and
2:22push it
2:23into our console.
2:26From there a language model and human
2:29researchers will triage the alert and
2:32conduct a deeper investigation.
2:36Okay, let me show you an example
2:38detection case
2:40detection case
2:42triggered inside the adapter built from
2:45call frame trace data.
2:47We reconstruct the cross side paths like
2:50this.
2:51And EOA interact with the same wallet
2:54supplying
2:56and supplying a malicious smart
2:58contract. This smart contract was
3:01deployed by the same EOA which is
3:06already tiny and labeled as Tornado Cash
3:09funded.
3:10The execution contains nested dedicate
3:14calls.
3:15Immediately this
3:17immediately
3:19automatically this replace the
3:21underlying implementation of the safe
3:24wallet and transfer full control to the
3:27attacker's smart contract.
3:30Then the stream matching engine it
3:31supports several pattern types equally
3:34including sequence patterns and absence
3:37patterns.
3:38Here is an example based on the layer
3:41zero exploit pattern.
3:45In this attack these two events have a
3:48gap of two blocks.
3:50My rule works like this. Once
3:53once we observe the package verified we
3:56will look backward to check whether a
3:59valid corresponding source
4:02package sending exist.
4:04If no matching source event can be
4:06found, the rule will trigger an alert.
4:11And moving on the investigation demo,
4:14after the stream matching engine push
4:18an alert to the console, what we can do?
4:23Okay, this is our console. You can see
4:25all the alerts has been listed here. And
4:28this is our data plane. You can click
4:31and see the details. And if you want to
4:34the
4:35language model agent to do the
4:37investigation, just click this and it
4:40will give you a
4:42analysis here.
4:43And also let me have a look at the
4:47control plane.
4:49We all the all the predefined patterns
4:53listed here.
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: 8ac5d2a12122de2c2e05bd3ca9107e22ca885cf6d9b27c269f7bfc203a08c753