Notifications
Clear all
Topic starter
18/08/2026 2:23 pm
TL;DR: Crypto exploits cost an estimated $3.4B in 2025, and Quantstamp argues the response is not more manual review but security agents that can standardise checks, verify findings, and scale alongside automated attackers. The real governance question is whether teams can evaluate agent output like infrastructure, with repeatable workflows and human validation.
NHIMG editorial — based on content published by Quantstamp: The Exploit Race
By the numbers:
- Prompt design alone can reduce false positives by over 60% in controlled evaluations.
- Crypto lost more than $240 million across 29 incidents in July 2026, up roughly 216% from June.
Questions worth separating out
Q: How should security teams use AI agents for vulnerability discovery without over-trusting them?
A: Treat AI agents as a repeatable screening layer, not as proof of security.
Q: Why do AI security agents need benchmark-driven governance?
A: Because performance varies by prompt design, context handling, and the evaluator itself.
Q: What breaks when vulnerability detection is done with raw prompting alone?
A: Raw prompting produces unstable results, inconsistent false positives, and invisible false negatives.
Practitioner guidance
- Standardise vulnerability prompts Use a fixed set of targeted prompts for common issue classes, then compare outputs across runs to identify noise, drift, and hidden false negatives.
- Separate discovery from verification Run one pass to flag candidate issues and a second pass to prove exploitability with traces, tests, or controlled reproduction.
- Benchmark against reproducible evidence Score agents on whether they identify the correct code path, preconditions, and failure mode, not just whether they resemble a known finding.
What's in the full article
Quantstamp's full blog post covers the operational detail this post intentionally leaves for the source:
- Benchmarking approaches for AI security agents in realistic smart-contract environments
- The specific workflow levers Quantstamp used to compare precision, recall, and verification quality
- Examples of multi-pass agent design across prompting, context handling, and tooling
- How Quantstamp thinks teams should integrate AI security checks into development cycles
👉 Read Quantstamp's analysis of the exploit race in web3 security →
Agentic security for web3 audits: what changes for teams now?
Explore further
18/08/2026 3:14 pm
Agentic security only becomes credible when it is governed like infrastructure. The article is right to reject raw prompting as a control model, because ad hoc AI output is not repeatable enough for security operations. In identity terms, this is the same lesson enterprises learned with service accounts and secrets: uncontrolled runtime authority creates ambiguity and risk. Practitioners should treat agent workflows as managed systems with defined scope, validation, and review.
A question worth separating out:
Q: How accountable are teams for AI-assisted security findings?
A: Teams remain accountable for the evidence behind any AI-assisted finding, just as they are for any other security control. If an agent cannot reproduce a claim, explain its basis, or show the relevant code path, the organisation should treat the output as advisory only and not as a decision record.
👉 Read our full editorial: The exploit race in web3 is turning security into automation
