Depthfirst Claims Cheaper AI Bug Hunting Than Anthropic Mythos — Why Crypto Infrastructure Should Pay Attention

May 12, 2026

On May 12, 2026, AI security startup Depthfirst said its autonomous vulnerability discovery model uncovered multiple high-severity Internet vulnerabilities that Anthropic’s Mythos program missed — and that it did so at roughly one-tenth of the cost. Depthfirst’s CEO Qasim Mithani framed the approach as task-specialized optimization: build a model architecture tightly tuned for vulnerability discovery so that “$1,000 of compute can do what $10,000 would otherwise buy.”

For the blockchain and crypto industry, the headline is not “who found more bugs.” The real story is that the economics of finding and weaponizing vulnerabilities are collapsing. That shift changes the threat model for everything that touches private keys: RPC gateways, staking dashboards, exchange hot-wallet pipelines, DeFi frontends, and even the browsers and operating systems used to sign transactions.

The New Security Reality: AI Lowers the Marginal Cost of Zero-Days

AI-driven vulnerability discovery is rapidly moving from “assistive” to “agentic”: models can search codebases, reason about exploitability, and iterate toward proof-of-concept conditions with minimal human input. Anthropic itself has described Mythos-class capability as a meaningful step-change in autonomous discovery and exploitation workflows (see Anthropic’s write-up on Mythos Preview under Project Glasswing: Claude Mythos Preview).

Depthfirst’s claim adds a second accelerant: cost efficiency. If a specialized model can produce comparable findings for ~10% of the spend, then:

Defensive teams can scan more code, more often.
Offensive actors can do the same — including targeting crypto-heavy infrastructure with thinner security margins.

This aligns with broader signals from threat intelligence. Google’s threat teams have warned that AI-assisted hacking is already appearing in the wild, including activity assessed to involve AI in the discovery and weaponization of a previously unknown “zero-day” style weakness (reporting on Google’s warning).

What Depthfirst Says It Found (And Why These Targets Matter to Web3)

Depthfirst’s announcement referenced several categories of issues across widely deployed software. Even without assuming every detail is fully public yet, the types of components mentioned are exactly the ones Web3 depends on:

1) NGINX: The Default Front Door for RPC, APIs, and Dashboards

Depthfirst says it found a critical flaw in NGINX, described as long-lived (dating back to 2008) and potentially impacting a large portion of the Internet, with a patch expected from the current maintainer (F5) soon.

In crypto, NGINX commonly sits in front of:

RPC endpoints (rate limiting, caching, routing, TLS termination)
Explorer and indexer APIs
Custody and signing service gateways (internal admin panels, service-to-service ingress)
Exchange and broker web layers

A single high-impact NGINX issue can become an “internet-wide scanning” event. For operators, the most practical move is to monitor upstream advisories continuously, because “patch Tuesday” thinking does not survive AI-scale discovery. Start with the upstream index: NGINX security advisories.

2) Linux: Validator Nodes, Indexers, and Signers Mostly Run Here

Depthfirst also referenced a serious Linux vulnerability with remote code execution potential, reportedly not fully remediated at the time of disclosure.

Whether the specific bug is in kernel space, common libraries, or a default service, the crypto takeaway is consistent:

Most node fleets run homogeneous Linux images.
Once a reliable exploit exists, attackers can pivot from “one box” to “many boxes,” especially where RPC nodes are exposed or where internal networks are flat.

If you maintain production infrastructure, keep your vulnerability intake tied to a coordinated disclosure process. The U.S. government’s guidance on coordinated vulnerability disclosure is a helpful baseline for programs and response workflows: CISA Coordinated Vulnerability Disclosure Program.

3) Chrome: The Wallet UX Layer Is a Browser Security Problem

Depthfirst’s statement also pointed to vulnerabilities in Google Chrome, with at least some already fixed by Google.

For crypto users, browsers are effectively part of the signing perimeter:

Wallet extensions and injected providers live here.
Many phishing attacks succeed without “breaking crypto” — they break the browser trust layer, then trick users into signing.

Operationally, the simplest control is still one of the strongest: keep browsers on stable, updated channels and enforce version compliance in enterprise settings. Track upstream fixes via the official channel: Chrome Releases.

4) FFmpeg: “Non-Crypto” Libraries Still End Up in Crypto Systems

Depthfirst also referenced issues in FFmpeg, an open-source multimedia framework. It may sound unrelated to blockchain — until you consider real deployments:

Support tooling that ingests user content (KYC videos, support attachments, marketing uploads)
Internal moderation systems
Notification pipelines that process media
Any “utility” microservice bundled into a larger platform image

FFmpeg’s own security guidance is here: FFmpeg Security.

The Web3 lesson: attackers do not care whether a library is “crypto-native.” They care whether it’s reachable, exploitable, and leads to credentials, secrets, or lateral movement toward signing systems.

Depthfirst’s “Open Defense Initiative”: Why Access Matters More Than Branding

Depthfirst also announced a new plan called the Open Defense Initiative, committing up to $5 million in access/credits to make its AI vulnerability detection tools available to enterprises and open-source developers.

If this model (and others like it) becomes widely available, we should expect two opposing effects at the same time:

Better security for open source, because maintainers and defenders can find bugs earlier.
More “bug-to-exploit” pressure, because discovery velocity rises faster than many projects can triage, patch, and ship releases.

This is not hypothetical. Even today, the vulnerability ecosystem is straining under volume. NIST has publicly discussed operational changes to handle record growth in CVE records and prioritization at scale (NIST NVD updates).

Why Crypto Is a High-Value Target in an AI-Accelerated World

Crypto systems concentrate value in ways that make them uniquely attractive when exploitation gets cheaper:

Private keys convert compromise into immediate, irreversible loss
A server RCE or supply-chain backdoor can lead directly to signing authority, hot-wallet drains, or malicious updates.
Open-source, composable stacks expand the attack surface
DeFi and infrastructure teams ship fast, reuse dependencies, and integrate third-party SDKs. AI makes it easier to search for weak links across that dependency graph.
Web3 is operationally “always-on”
Nodes, indexers, relayers, bridges, and market-making systems can’t simply pause for a long patch window without real financial impact — which is exactly what attackers exploit.
Attackers are already adapting
Beyond vendor announcements, threat intelligence reporting now points to criminals experimenting with AI to discover and operationalize previously unknown weaknesses (coverage of Google’s assessment).

A Practical Defense Checklist for Web3 Teams

If you operate crypto infrastructure (wallet services, exchanges, RPC providers, protocols, or dApps), consider the following priorities as “AI-era basics”:

1) Treat Internet-facing components as disposable

Use immutable images, rapid rebuilds, and blue/green deploys.
Assume compromise and design for fast rotation.

2) Reduce blast radius around signing authority

Isolate signers from general compute.
Strictly segment networks; eliminate inbound paths to signing hosts.
Require multiple controls for production signing (policy + human review + rate limits).

3) Move from “scan everything” to “validate exploitability”

AI-driven tools will flood queues if your process can’t validate reachability and exploit paths. Make sure your pipeline can answer:

Is it reachable in our deployment?
Is it exploitable with our config?
Does it touch secrets, auth, or signing?

4) Harden the supply chain

Pin dependencies and verify integrity.
Use reproducible builds where possible.
Sign releases and verify signatures in deployment.
Watch for typosquatting in build-time packages (especially in CI).

5) Upgrade disclosure and incident response muscle

If AI can find issues overnight, your organization needs to:

Triage quickly
Patch safely
Communicate clearly
Coordinate upstream when the bug is in shared infrastructure

CISA’s disclosure guidance is a good starting point for program structure and expectations: CISA vulnerability disclosure resources.

What This Means for Everyday Crypto Users

Even if you don’t run servers, AI-accelerated vulnerability discovery still affects you because your transactions touch:

Browsers (phishing + browser bugs)
Operating systems (credential theft, clipboard hijacking, session token theft)
Websites and RPC endpoints (malicious routing, injected content)

Two simple behaviors matter more than ever:

Update your browser and OS promptly
Track Chrome fixes via the official release channel: Chrome Releases.
Keep private keys isolated from the internet-facing attack surface
This is where a hardware wallet meaningfully reduces risk: even if your computer is exposed to a browser exploit, a well-designed signing flow forces on-device confirmation and prevents raw private keys from ever touching the host.

Where OneKey Fits: Defense-in-Depth for the AI Era

AI will make vulnerability discovery faster on both sides of the fight. In that environment, the most resilient crypto security posture is layered:

Infrastructure hygiene (patching, segmentation, least privilege)
Operational controls (signing policy, monitoring, incident response)
User-side key isolation (hardware-backed signing)

For users focused on self-custody, OneKey’s core value proposition maps directly to this reality: keeping private keys off general-purpose devices and requiring explicit confirmation during signing. As the browser, OS, and web server layers become more volatile under AI-scale discovery, isolating keys becomes less of a “best practice” and more of a baseline expectation for serious crypto security.