Web3 Pentest Findings in DeFi Backends, Keepers and Cloud

Updated November 2026 5 min read

Quick answer. In 331 offchain findings from 19 published Guardian reports, the largest groups are offchain math that drifts from the chain, lost or raced messages, unvalidated inputs and missing rate limits. 27 of them are Critical or High.

Watch: Web3 Pentest Findings in DeFi Backends, Keepers and Cloud 1:43

Video transcript

Offchain pentest findings are rarely published. Here is what Guardian's reviews found in DeFi backends, APIs, keepers and cloud.

By Guardian's count, offchain gaps accounted for seven hundred twenty nine million dollars of the losses in 2026's ten largest exploits.

The set is three hundred thirty one findings from nineteen public reports, plus three case studies. It is small, so read it as Guardian's findings, not industry statistics.

The biggest groups are accounting that drifts from the chain, messages that get lost or raced, unvalidated inputs, and missing rate limits.

Unlike smart contracts, the worst offchain bugs are rarely clever. A service computes a different number, drops a message, or accepts unlimited work.

A Critical in Ginza Gaming's backend let users submit their own audit result and skip review. At Synthetix, a failed withdrawal still returned success, leaving funds locked.

GMX's keeper review found eight Highs, mostly where its view disagreed with the contracts. A Synthetix keeper sat behind a public API Gateway with no authorizer, and its logs captured private keys.

Fix in order. Make offchain math match the chain, and test them together. Never acknowledge a message before its work is durable. Bound every request per real client. Then sign whole payloads and split keys.

Guardian's offchain team is led by Carlos Polop, creator of HackTricks. Read the full report, or talk to Guardian about the systems around your contracts.

Web3 pentest findings are rarely published. Guardian's own portfolio shows the gap: 332 published reports are filed as onchain work and only five as offchain.

The gap matters. At the launch of Vanguard, Guardian's count put $729M of the $769M lost in the ten largest exploits of 2026 down to offchain gaps.

This report collects what Guardian has published about offchain systems. It is a small, concentrated set, so read it as Guardian's findings, not industry statistics.

It is the third of three reports on Guardian's public data, after What 360 Critical Bugs Taught Us and Smart Contract Upgrade Bugs.

Dataset Scope

The sources are Guardian's published reports as of November 4, 2026 (337 reports, 10,086 findings) and its offchain case studies.

331 findings have published details: 70 in two reports filed as offchain, and 261 inside 17 reports filed as onchain.

Three case studies add severity counts only: Avantis (two Criticals, six Highs), GMX's keeper (eight Highs) and USDT0's Gasolina signer (no Critical or High).

How we selected and counted:

Offchain reports. Two of the five filed as offchain; the other three review contract code.
Onchain reports. Findings in offchain files or services, or labeled Keeper. Excluding Global-only findings makes this a floor.
Case studies. Their finding details remain private, so they are not in the theme counts.
Themes. Keyword matches on finding titles. A finding can match several.

Limitations. The set is small and concentrated. One report, Synthetix Withdrawal Flow, holds 17 of the 27 Critical and High findings, and its 83 findings have no final status recorded.

There is no published browser extension or SDK pentest in the set.

What the Findings Were About

Theme (by title)	Findings	Critical or High	Reports
Prices, precision and accounting	35	7	10
Message handling, races and lost or stale state	34	6	7
Untrusted inputs and external responses	29	2	7
Rate limits, quotas and resource exhaustion	22	5	6
Signatures, replay and payload binding	19	2	6
Authentication, sessions and privileged actions	15	0	6
Secrets, keys and credentials	15	1	4
Cloud and infrastructure configuration	13	0	4

The pattern differs from smart contracts. The highest-severity offchain findings are rarely about a clever exploit.

More often, a service computes a different number from the chain, drops a message or accepts unlimited work from anyone.

Backends, Keepers and Cloud

Trading backends and APIs

Synthetix's Withdrawal Flow review covered a Go backend and returned 17 Highs. Several show how a backend can lock user funds without any attacker.

When submitting a withdrawal failed, for example during an RPC outage, the handler still returned success. The queue discarded the request while the user's pending amount stayed locked (H-06).

Others are classic API abuse: no rate limit on withdrawal requests, server timeouts never applied, and request bodies read without a size limit.

Trade actions were also signed without fields such as withdrawal amount and destination, so a tampered request still verified (H-15).

In Ethereal's Orderbook DEX, Round 2, a WebSocket rate limiter logged an exceeded quota but still let the request through (H-01).

Rate limits keyed to the load balancer's address, not the client's, turned up there and in Longshot's Rust API. Longshot also lacked step-up 2FA on admin and vault routes that move funds.

Ginza Gaming's backend (C-05) holds a Critical. A withdrawal request accepted a user-supplied audit result, so a user could mark their own withdrawal as passed and skip review.

In the Avantis V2 backend, Guardian found two Criticals and six Highs. Both Criticals were fixed while the review was still running.

Keepers and automation

Keepers hold funded keys and act unattended. In this set, their serious bugs were mostly wrong numbers rather than broken access.

GMX keeper review: 55 findings and eight Highs, all eight resolved, mostly where the keeper disagreed with the contracts.
GMX operations script: double counted open interest when long and short collateral are the same token (High).
Bounce Tech automation: used position size where it needed margin, corrupting the target leverage calculation (High).
Synthetix LP vault keeper: swapped every reward token it received, including unconfigured ones, creating sandwich and slippage risk (High).

Cloud and secrets

The clearest cloud findings come from the Synthetix LP vault keeper on AWS. A public API Gateway endpoint had no authorizer, so anyone with the URL could invoke the keeper's Lambda.

Per the case study, request logging wrote full payloads, including private keys, to CloudWatch. Secrets and an RPC key were hard-coded in source.

Other reports add the same families. Bounce Tech's Terraform exposed a Kubernetes control plane to 0.0.0.0/0, and an account script printed private keys to the console.

In Synthetix's withdrawal backend, one private key held the relayer, watcher and teller roles (H-14). So the watcher votes meant to check withdrawals came from the key that requested and paid them.

Fix Priorities

The order follows what produced the most Critical and High findings in this set.

Make offchain math match onchain math, and test them against each other (7 Critical or High).
Never acknowledge a message before its work is durable (6 Critical or High).
Bound everything, per real client: rate limits, timeouts and body size limits (5 Critical or High).
Sign the whole payload, with a nonce and an expiry.
Treat every client-supplied field as hostile, including a review status (the Ginza Gaming Critical).
Split keys by role, and keep secrets out of code and logs.
Authenticate every endpoint that triggers value movement, with step-up 2FA for privileged actions.
Lock down cloud defaults: control-plane access, security groups, WAF and log contents.

How Guardian Tests Offchain Systems

Guardian's offchain services are led by Carlos Polop, Head of Offchain Services, who created HackTricks and the PEASS privilege escalation suite.

The team runs web app, API, infrastructure, browser extension and offchain automation pentests and SDK audits. It works beside the researchers reviewing the contracts.

For onchain teams, Vanguard returns half the package cost as credits toward offchain assessments. Its six months of update review cover up to 10,000 SLOC of offchain code.

For scope and price, see how much a web3 pentest costs and the best web3 pentesting firms.

Frequently Asked Questions

What vulnerabilities do pentests find in DeFi web apps and backends?

In Guardian's published offchain findings, the most frequent are offchain math that drifts from the chain, lost or raced messages, unvalidated inputs and missing rate limits. Replayable or partly signed messages, exposed secrets and cloud defaults follow.

How severe are offchain findings compared with smart contract findings?

In the 331 published offchain findings with details, 27 (8.2%) are Critical or High, against 12.5% in reports filed as onchain. The scopes differ, and offchain Criticals do occur: the Avantis V2 backend review found two.

Are keepers and bots in scope for a web3 pentest?

They should be. Guardian's GMX keeper review returned 55 findings, eight of them High, mostly where the keeper's view of a position disagreed with the contracts.

What should a DeFi team fix first after an offchain pentest?

Start with anything that moves value: offchain math that disagrees with the contracts, messages acknowledged too early, replayable payloads, exposed keys and unauthenticated endpoints. Then add rate limits, timeouts and cloud hardening.

Sources

Pentest the systems around your contracts

Guardian's offchain team tests the backends, keepers, APIs and cloud accounts that decide what your contracts are asked to do.

Talk to Guardian Read the Avantis offchain case study