Hacking SAML with Claude Code news update

In an empirical security research report published by oblique.security's report, security researcher Niels Provos's methodology and Anthropic's Cyber Verification Program were leveraged to evaluate SAML protocol implementations using Claude Opus. Following acceptance into Anthropic's program to remove standard safety guardrails, the researcher constructed an automated hacking harness to generate exploits across every accessible SAML implementation. Over a month of testing, the research pipeline uncovered full authentication bypasses in major projects, dozen auxiliary component signature bypasses, and unauthenticated denial-of-service vulnerabilities across multiple programming language ecosystems.

This news update examines the architectural design of the automated SAML exploitation harness built around Claude Opus, the specific vulnerabilities discovered across target frameworks, and the defensive recommendations for development teams handling SAML assertion processing. The analysis details how model-driven vulnerability research operates at scale, evaluating the performance metrics of AI-generated exploits, emergency patching responses across target maintainers, and instructions for integrating security analysis harnesses within enterprise workflows.

What shipped in Hacking SAML with Claude Code news update

The research release details an open-source automated vulnerability discovery harness engineered around Anthropic's Claude Opus model, hosted on GitHub under oblique-security/saml-research. Operating on an upgraded Claude Max 20x plan, the system breaks technical security auditing into two distinct operational phases to scale context management: a gadget discovery phase that identifies anomalous parsing primitives in XML libraries, followed by a findings confirmation phase that synthesizes identified gadgets into verified end-to-end exploits. Intermediate data structures, gadget definitions, and scope evaluation rules are structured as JSONL records to prevent token waste on out-of-scope code paths.

Rather than relying on explicit pattern-matching instructions for historical vulnerabilities, the harness feeds high-level threat models and general guidance to Claude Opus, allowing the model to independently explore attack surfaces. The automated pipeline discovered four zero-day authentication bypasses: an XML comment injection flaw in NameID handling for Authentik tracked under CVE-2026-57580, signature wrapping vulnerabilities on Response messages in PHP's litesaml/lightsaml package under CVE-2026-63182, signature wrapping in OneUptime issue 2949, and signature wrapping in Java's saml-client project under issue 149. Additionally, twelve auxiliary signature bypasses were confirmed in non-Response messages such as AuthnRequest, AttributeQuery, and LogoutRequest, including an information disclosure vulnerability in TypeScript's samlify project.

What improved in Hacking SAML with Claude Code news update

Automated security research efficacy improved through structured multi-agent coordination, intermediate data storage, and strict scoping rules that optimize LLM token usage during exploit generation. In testing against Node's xml-crypto library, the harness detected canonicalization flattening (gadget g-0005) where processing instructions were stripped to bare strings, which the confirmation phase automatically escalated into a functional email truncation authentication bypass. Memory management analysis during assertion validation led to accepted upstream patches in Go's xmldsig package to fix quadratic memory allocation during signature verification, while identifying active unauthenticated out-of-memory denial-of-service vectors in Node's xmldom library and Python packages utilizing libxmlsec1 XSLT transformations.

Maintainer remediation responses varied across ecosystems following vulnerability submission. For the Authentik vulnerability, eight independent security researchers submitted reports concurrently using AI-assisted analysis tools. In the OneUptime repository, public disclosures triggered rapid automated AI-driven pull requests (#2949, #2981, and #2988) to resolve signature bypasses via signed error responses and processing instruction manipulations.

What you gain from Hacking SAML with Claude Code news update

Security engineering teams gain empirical evidence regarding the speed and thoroughness of model-driven protocol auditing, demonstrating that modern LLMs can discover complex logic bugs without explicit attack templates when provided appropriate harness primitives. Organizations deploying SAML authentication implementations obtain actionable vulnerability data across Go, Python, Node, PHP, and Java ecosystems, enabling immediate patching against unauthenticated memory exhaustion attacks and assertion spoofing. Developers receive concrete guidance urging the deprecation of custom SAML implementations in favor of standardized OpenID Connect protocols or hardened, vendor-maintained library APIs.

Enterprise security operations teams gain a reusable architectural blueprint for enterprise vulnerability discovery, utilizing JSONL structured pipelines to filter false positives prior to triggering execution-heavy exploit validation steps. The research highlights critical operational realities regarding open-source security maintainability, establishing that automated slop reports and high-volume AI vulnerability disclosures require formalized, automated triage mechanisms to prevent maintainer burnout.

How to get Hacking SAML with Claude Code news update

To deploy Claude Code and Anthropic models for security analysis and protocol research, update the CLI tool directly through standard package management channels. Execute the update command to ensure access to the latest Opus model capabilities:

npm install -g @anthropic-ai/claude-code

To switch models within an active interactive Claude Code terminal session to Claude Opus or specific research snapshots, execute the model configuration command:

/model claude-opus-3-5

To launch a security analysis session with a pinned model selection from the system shell command line, use the model flag directly:

claude --model opus

To establish a persistent default model selection across all CLI executions, configure your local environment settings file or export the primary model environment variable:

export CLAUDE_DEFAULT_MODEL="claude-opus-3-5"

What to watch after Hacking SAML with Claude Code news update

Security teams must track ongoing patching efforts across Python and Node SAML dependencies affected by unauthenticated XML denial-of-service vectors. Python implementations relying on libxmlsec1 require immediate monitoring for XSLT transform filter patches to block recursive template expansion attacks. Development teams using Node libraries dependent on xmldom should monitor security advisories for the resolution of private memory allocation reports.

Furthermore, enterprise open-source maintainers are expected to implement updated contribution guidelines and automated verification steps to handle the rising volume of AI-generated security disclosures. As multi-agent hacking harnesses become widely accessible, protocol maintainers must prepare for concurrent multi-reporter disclosures on complex XML signature wrapping vulnerabilities across public repositories.