ESL Security Response · Updated August 2026

An Answer to Tannenbaum's Concern

How we secure AI-generated HTML presentations before they reach the client's browser

9Security checks

3Defense layers

74+Files protected

The Concern

What Tannenbaum Raised

“Everything sent to the customer's browser is visible!”— Amit Tannenbaum, Data Analyst Team Lead, LinkedIn post · July 2026

Tannenbaum warned that AI-generated HTML can expose hidden JavaScript, embedded prompts, API keys, and sensitive business logic.

01 SAST

02 HTML Validator

03 Secrets Scanning

04 Privacy Scan

05 Business Logic

06 DAST

07 Network Inspection

08 Content Review

09 SHA-256

Our Response

We Ran the Assessment on Ourselves

Using the Claude Mythos Security workflow, we assessed 74 AI-generated HTML presentations at zuwasi.github.io/Public-html-pages.

✓ What We Did NOT Find

No leaked AI prompts
No real API keys
No data exfiltration
No malicious JS
No eval()
No prompt injection

⚠ What We DID Find

2 ERROR — plaintext passwords
7 WARN — tracking, SRI, YouTube
325 INFO — missing CSP, unsafe links
Cosmetic password gates
CDN without integrity
Visitor metadata sent

The Gap

General-Purpose Scanners Are Not Enough

In the original 74-file assessment, dependency, secrets, and SAST scans found no application vulnerabilities. The HTML-specific review still found the content and privacy issues below.

Finding	Why general scanners miss it	Our Checker ID
Password 86999 in comment	5-digit number, no secret pattern	SEC-01
Password rty768 in JS	Short string in comparison	SEC-02
Abacus tracking fetch()	Intentional analytics, not vuln	SEC-03
CDN script without SRI	Supply-chain best practice	SEC-04
Missing CSP on 73 files	Defense-in-depth gap	SEC-05
YouTube tracking embeds	Privacy concern	SEC-06
Leaked AI prompts	Content review, not SAST	SEC-07
Unsafe target=_blank	HTML best practice	SEC-08
Unversioned CDN script	Version pinning, not CVE	SEC-09

The Solution

A 3-Layer Defense Pipeline

Endor catches what it can, our custom checker catches the rest, and git hooks + CI enforce both.

1. Endor Labs

Dependency & Known-Secret Scan

Catches API keys, CVEs, SBOM.

↓

2. Custom Pre-Publication Checker

9 HTML-Specific Checks

Passwords, tracking, SRI, CSP, prompts.

↓

3. Git Hooks + GitHub Actions

Automated Enforcement

Pre-commit, pre-push, CI gate.

Layer 1

Endor Labs — The Foundation

✓ What Endor Catches

Real API keys (sk-*, AKIA*, ghp_*)
Dependency CVEs (npm, pip, maven)
Known secret formats (AWS, GitHub, private keys)

✗ What Endor Misses

Short passwords (86999, rty768 — no entropy match)
Intentional tracking (analytics fetch ≠ vulnerability)
Content-level issues (missing CSP, leaked prompts, SRI)

THE BLIND SPOT

Why Not Just Ask AI to Review Its Own HTML?

"If Claude generated the HTML, asking Claude to review it is like asking a developer to review their own code — same blind spots in both roles."The Self-Attribution Problem

✗ AI Reviewing AI

Shared blind spots — AI doesn't flag what it doesn't consider dangerous
Pattern reinforcement — "That's just context I wrote"
False confidence — "I reviewed it, it's clean"
Adversarial risk — prompt injection can fool the reviewer too

✓ Pure Regex Scanner

No AI bias — doesn't "think", just matches strings
No context reasoning — flags sk-proj-abc123 every time
Caught what AI missed — password 86999, prompt leakage, tracking
Zero LLM tokens — runs in milliseconds, not minutes

Diverse tools catch what each other miss. That's defense in depth.

Layer 2

The Pre-Publication Checker

A dependency-free Python CLI and local browser GUI perform nine deterministic HTML-specific checks.

python -m pip install "git+https://github.com/zuwasi/html-report-security-checker.git@6d21c5c"
html-security-checker report.html
html-security-checker reports/ --fail-on-warning
html-security-checker --gui

ID	Tannenbaum mapping	General scanner	ID	Tannenbaum mapping	General scanner
SEC-01	Secrets	May miss	SEC-06	Privacy	May miss
SEC-02	Business logic	May miss	SEC-07	Content	May miss
SEC-03	Network / tracking	May miss	SEC-08	HTML safety	May miss
SEC-04	SRI / supply chain	May miss	SEC-09	Version pinning	May miss
SEC-05	CSP	May miss

Layer 3

Automated Enforcement

Three enforcement points. No manual step required.

Pre-Commit Hook

Runs on git commit, scans staged .html files, stays fast, blocks on error.

Pre-Push Hook

Runs on git push, full repository scan with -FailOnWarning, blocks on warning.

GitHub Actions CI

Runs on PR/push to main, blocks merge, and generates the SHA-256 manifest.

Result

Tannenbaum's Checklist — Covered by the Pipeline

#	Checklist item	Endor	Custom checker	CI/CD
1	SAST	PASS	PASS	Auto
2	HTML Validator	—	MANUAL	Review
3	Secrets Scanning	PASS	PASS	Auto
4	Privacy Scan	—	PASS	Auto
5	Business Logic	—	PASS	Auto
6	DAST	—	MANUAL	Review
7	Network Inspection	—	PASS	Auto
8	Content Review	—	PASS	Auto
9	SHA-256	—	PASS	Auto

7 of 9 automated · HTML validation and DAST remain explicit human-review activities.

Verification

What the HTML-Specific Review Found

2ERRORS

7WARNINGS

325INFO

Key point

All 334 content, privacy, and browser-safety findings required the HTML-specific layer. This complements—not replaces—SAST, secrets, dependency, and CVE scanning.

Tannenbaum Item #9

SHA-256 Artifact Integrity

CI generates a manifest automatically on every push to main.

# SHA256-MANIFEST.txt
73e4c9...  ./index.html
a510bd...  ./answer-to-tannenbaum-concern.html
19d0a7...  ./security-review/index.html

How it works

The security-check.yml workflow computes SHA-256 for every HTML file and uploads the manifest as a GitHub artifact retained for 90 days.

Practical

How to Use the Pipeline

1. Install

git clone https://github.com/zuwasi/html-report-security-checker.git
cd html-report-security-checker
python -m pip install .

2. Scan

html-security-checker report.html
html-security-checker reports/ --fail-on-warning
html-security-checker --gui

3. Enforce

./install.sh
git add report.html
git commit
git push

Hooks and CI prevent unsafe reports from being published accidentally.

August 2026 Hardening Update

We Reassessed the Checker Itself

Application

51 automated tests passed
Zero runtime dependencies
Wheel build and Bandit scan passed

Local GUI

Same-origin POST enforcement
JSON-only API requests
Cross-site file writes blocked

Git & CI

Hooks no longer import from an untrusted repo root
Bash hooks forced to LF line endings
Workflow permissions reduced

CVE Review

CVE-2024-42471 fixed in download-artifact 4.1.3+
Exact v4.3.0 commit pinned
All GitHub Actions pinned by SHA

The Answer

Tannenbaum Was Right — Build a Safety Gate

Defense in depth

Dependency + CVE scanning
  → Deterministic HTML checks
  → Git hooks + CI
  → Human review
  → Publish with evidence

Use it. Improve it. Challenge it.

HTML Report Security Checker is MIT licensed. Contributions, new detection rules, false-positive reports, stars, and practical feedback are welcome.

GitHub repository ↗Daniel Liezrowice on LinkedIn ↗Contact ESL ↗