Three significant stories closed out the week of July 28 — OpenAI quietly open-sourced a security scanning tool, Amazon attribution linked major npm supply chain attacks to a North Korean threat actor, and Anthropic published findings from its Mythos AI model that include novel cryptographic attacks affecting widely deployed algorithms.
OpenAI Releases Codex Security CLI (Open Source)
On July 29, OpenAI pushed @openai/codex-security (v0.1.1, Apache 2.0) to GitHub before officially announcing it — the Hacker News community found it first. OpenAI later acknowledged the early discovery: "quietly released the open-source Codex Security CLI, but Hacker News found it before we had a chance to share it here…"
The CLI scans source code repositories for security vulnerabilities, performs security-focused code reviews, provides remediation recommendations, tracks findings across runs to verify fixes, and integrates into CI/CD pipelines. It accumulated approximately 2,600 GitHub stars within a single day of going public.
The catch: while the CLI's source code is open, the underlying vulnerability scanner engine remains gated to approved enterprise customers. Organizations without that access get the plumbing without the engine — limiting the tool's immediate utility for the broader developer community. The release is widely seen as a direct competitive response to Anthropic's Claude security tooling.
AWS Links NPM Supply Chain Attacks to North Korea's Sapphire Sleet
Amazon Threat Intelligence published findings on July 30 attributing a series of npm package supply chain attacks to Sapphire Sleet (also tracked as STARDUST CHOLLIMA, BlueNoroff, CageyChameleon, and Alluring Pisces) — a North Korean threat actor historically associated with cryptocurrency theft and developer-targeting social engineering campaigns.
Packages Compromised
| Package | Date | Weekly Downloads |
|---|---|---|
| typo-crypto | March 2025 | Low (assessed as test run) |
| debug | September 2025 | Hundreds of millions |
| chalk | September 2025 | Hundreds of millions |
| axios | March 2026 | 100+ million |
Attack Methodology
Attackers socially engineered the maintainers of legitimate, high-download packages to gain account access, then injected malicious postinstall scripts that execute automatically during npm install. The payloads are environment-aware — detecting Windows, macOS, or Linux and downloading a tailored second-stage payload accordingly.
The typo-crypto payload activated only after receiving a specific hash input before downloading its second stage, a fingerprinting technique designed to avoid sandbox detection. Fragmented payloads were used to evade static analysis tools.
Scale: Research cited by AWS found approximately 1 in 10 cloud environments was affected within a two-hour window following the debug/chalk compromise. Amazon CISO CJ Moses stated: "By compromising a small number of highly popular packages, the group gains potential access to thousands of downstream environments simultaneously."
This is the first public attribution of these specific compromises to Sapphire Sleet. Attribution confidence is assessed as medium.
Anthropic Mythos AI Finds Cryptographic Weaknesses
Anthropic has published the latest findings from Mythos, its restricted frontier AI model made available to select tech firms and infrastructure providers under Project Glasswing for defensive cybersecurity research. The results span both vulnerability discovery at scale and original cryptographic research.
OSS Vulnerability Discovery
Mythos scanned 1,000 open-source projects and identified 23,000 potential vulnerabilities. Of 1,900 reviewed by external security firms, 1,726 were confirmed — including over 1,000 rated high or critical severity. Anthropic estimates approximately 3,900 critical and high-severity vulnerabilities will ultimately be confirmed from the full scan.
Notable findings include:
- 271 vulnerabilities in Firefox
- CVE-2026-4747: A 17-year-old FreeBSD NFS remote code execution bug that an unauthenticated attacker could use to gain complete root access — identified and exploited fully autonomously by Mythos
- A 27-year-old bug in OpenBSD
As of May 22, 2026: 1,596 vulnerabilities disclosed across 281 projects, with 97 patched.
Cryptographic Research
Mythos developed two cryptographic attacks of note:
- An improved attack on HAWK, a post-quantum signature scheme that is a NIST third-round candidate for additional post-quantum signatures
- A new attack on a reduced version of AES, the world's most widely used symmetric encryption standard — Anthropic states this has no immediate impact on deployed systems
Additionally, Mythos exploited vulnerabilities in wolfSSL (used by billions of devices) to forge TLS certificates, enabling an attacker to host a convincing fraudulent site that appears fully legitimate to end users.
Each cryptographic attack was developed in a multi-agent system at an API cost of approximately $100,000, with human researchers primarily handling project management and result verification. The SecurityWeek roundup framed these findings in terms of the cryptocurrency industry's exposure — virtually every DeFi protocol, wallet, exchange, and blockchain node client depends on the open-source libraries, operating systems, and networking stacks within scope of Mythos' scan.
The convergence of AI-accelerated vulnerability research, sophisticated state-sponsored supply chain operations, and the commercialization of security tooling marks a notable shift in the threat landscape heading into Black Hat USA 2026.