RESEARCH


← back

Pre-Install Analysis for Supply Chain Defense

April 2026security, npm

npm packages execute arbitrary code at install time, before anyone reviews a single line. Every existing defense checks a vulnerability database and catches nothing that hasn't already been reported. Supply chain attacks are zero-day by nature.

The fix is a pre-install gate: intercept the package manager, dry-run resolve, download, and run six parallel analyzers against every dependency change before it touches disk. No vulnerability database required. It catches install-script injection, typosquatting, dependency confusion, maintainer compromise, obfuscated payloads, and manifest confusion in one pass.

331 tests. 17 end-to-end against crafted attack packages, one for every major supply chain vector. Zero false negatives. Zero false positives across seven clean-package edge cases. Full pipeline runs in under 3 seconds on 500-dependency projects.

Every gate decision is HMAC-chained into a tamper-resistant audit log with cryptographic proof the record hasn't been modified after the fact. No trust required.

Enter PIN to access the full paper:

Don't have a PIN? Reach out at hey@ak68a.co.

Evaluating Agent Reliability in Financial Tool Use

March 2026fintech, AI agents

AI agents using financial tools fail 75% of the time out of the box. Failures span malformed arguments, broken data chaining, and non-deterministic tool selection, and they're model-independent. Validated across Claude Sonnet 4 and GPT-4o: both fail identically.

The fix is five layers stacked: Zod schemas enforce argument shapes, a retry loop catches transient failures, output labels let the model chain data across calls, schema-rich prompts ground tool selection, and a correction layer catches wrong-tool picks. Remove any one and the system regresses.

Single-run evals give false confidence. The eval harness runs each workflow multiple times and tracks pass rates per layer, so non-deterministic failures and regressions surface immediately.

Built on the Agent Commerce Kit (ACK) protocol with real cryptographic operations: W3C Verifiable Credentials, DID resolution, Ed25519 signing, and payment request/receipt issuance over JSON-RPC. No mocks, no stubs.

Enter PIN to access the full paper:

Don't have a PIN? Reach out at hey@ak68a.co.