Independent reproduction will determine whether Claude Code can autonomously discover and exploit consequential SAML implementation flaws in realistic applications.

state: expiredheat: lowuncertainty: highknownscott: mediumagentic-security coding-agents claude-codeEric ChiangOblique SecurityAnthropic

What is this?

The case concerns a reported experiment by Eric Chiang of Oblique Security using Anthropic’s Claude Code to search SAML implementations for exploitable security flaws. The supplied search results do not independently document that specific experiment or its findings, but broader testing shows Claude Code can find real vulnerabilities while producing high false-positive rates; source-only analysis also struggles to establish exploitability in deployed systems. Independent reproduction would therefore need to verify both the claimed SAML flaws and Claude Code’s ability to discover and exploit them autonomously under realistic conditions.

Why it matters to Scott

Scott already holds the case’s central position in Security Reviewer Method: agent-found vulnerabilities remain conditional until dangerous primitives are closed into reachable, independently checkable exploit paths. Reproducing the SAML work would directly exercise his active security-review harness and Claude Code practice, but the supplied evidence does not yet establish a new result; the radar also already tracks closely related Claude Code security validation in “Claude Security plugin beta.”

ip:source.security-reviewer-method-ebookip:concept.claim-bounded-adversarial-verificationip:concept.mechanically-different-verifiersdev:project.wordpress-security-reviewdev:technology.claude-coderadar:claude-security-plugin-betaradar:exploitgym-agent-exploitation-validationradar:concept.coding-agent-securityradar:concept.agentic-security

queries asked of Scott's wikis

autonomous coding agents for vulnerability research
agent harnesses for exploit validation and false-positive reduction
coding-agent evaluation on realistic end-to-end tasks
AI security research reproducibility and independent verification
agentic testing of authentication and identity protocols
source analysis versus runtime exploitability

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	Hacking SAML with Claude Code	ericchiang	1	0
🟧 echo.blog ⭐	This is the original account of Eric Chiang’s SAML research using Claude Code. He says he “attempted to hack every SAML implementation I cou	Oblique	—	—

Interpretation history

2026-08-22T00:23:19Z
After the monitoring horizon, the account still has no independent reproduction, named affected implementation, concrete flaw, or reachable exploit path. The episode has faded as a standalone signal and should be reopened only if a reproducible technical artifact emerges.
2026-08-19T23:39:02Z
No independent reproduction, concrete vulnerability, affected implementation, or reachable exploit path has emerged; this remains an unvalidated primary-account claim rather than a developing result. The unchanged observation adds no new meaning, so the case cools while awaiting substantive technical evidence.
2026-08-19T23:32:13Z
grounded: known/medium — Scott already holds the case’s central position in Security Reviewer Method: agent-found vulnerabilities remain conditional until dangerous primitives are close
2026-08-19T23:29:33Z
origin walked (codex/luna, conf 0.98): anchor hn.story.49368038 -> echo.blog.73c8f84e29 by Oblique
2026-08-19T23:28:24Z
case created — The original security write-up presents a concrete coding-agent vulnerability-research episode with transferable control and evaluation lessons.

Decision trace

08-22 10:23expireAfter the monitoring horizon, the account still has no independent reproduction, named affected implementation, concrete flaw, or reachable exploit path. The episode has faded as a standalone signal a
08-22 10:23alert_silentThe only change is negligible engagement on the same unvalidated primary account; no consequential technical delta would be lost by waiting.
08-22 10:23alert_routeThe only change is negligible engagement on the same unvalidated primary account; no consequential technical delta would be lost by waiting.
08-20 09:39repriceNo independent reproduction, concrete vulnerability, affected implementation, or reachable exploit path has emerged; this remains an unvalidated primary-account claim rather than a developing result.
08-20 09:39alert_silentThe new delta is only a legacy-state re-evaluation with no fresh evidence or engagement; nothing consequential would be lost by waiting for a normal briefing or a reproducible finding.
08-20 09:39alert_routeThe new delta is only a legacy-state re-evaluation with no fresh evidence or engagement; nothing consequential would be lost by waiting for a normal briefing or a reproducible finding.
08-20 09:37alert_silentThe primary account establishes that Eric Chiang used Claude Code to investigate multiple SAML implementations, but the supplied evidence does not identify a concrete vulnerability, affected implement
08-20 09:37surface_candidateThe primary account establishes that Eric Chiang used Claude Code to investigate multiple SAML implementations, but the supplied evidence does not identify a concrete vulnerability, affected implement
08-20 09:37alert_routeThe primary account establishes that Eric Chiang used Claude Code to investigate multiple SAML implementations, but the supplied evidence does not identify a concrete vulnerability, affected implement
08-20 09:32groundScott already holds the case’s central position in Security Reviewer Method: agent-found vulnerabilities remain conditional until dangerous primitives are closed into reachable, independently checkabl
08-20 09:29promote_anchororigin walk conf 0.98
08-20 09:28createThe original security write-up presents a concrete coding-agent vulnerability-research episode with transferable control and evaluation lessons.