False positive cybersecurity risk flag for benign Hertz price comparison/browser automation task

Summary

Codex incorrectly flagged a benign car-rental price comparison task as a possible cybersecurity risk.

The task was to use a real browser to check publicly visible Hertz.com rental prices around UCSD / La Jolla / San Diego across different dates, locations, and public Hertz discount / coupon parameters. The prompt explicitly prohibited submitting any reservation, entering payment information, logging into accounts, using real AAA membership numbers, adding insurance or paid add-ons, or bypassing website protections.

This appears to be a false positive triggered by normal browser automation / price-checking language.

Correlation ID

168328d5-8583-488c-9b7f-72dcf4281840

The in-app feedback was already submitted with current Codex session logs included.

Environment

• App: Codex desktop app
• Codex version: 26.506.31421 (2620)
• OS: macOS Tahoe 26.4.1
• Hardware: MacBook Pro, 16-inch, Apple M5 Max, 128 GB memory
• Browser task: Hertz.com price comparison using Chrome / browser automation
• Date: 2026-05-14 Pacific Time

Actual behavior

During the task, Codex showed:

This content was flagged for possible cybersecurity risk.

The chat also displayed:

This chat was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request or submit /feedback.

After submitting feedback, the app returned the correlation ID:

168328d5-8583-488c-9b7f-72dcf4281840

This interrupted a normal consumer price-comparison workflow.

Expected behavior

Codex should allow this type of benign browser task.

The prompt was not asking Codex to:

• bypass authentication
• bypass bot detection or CAPTCHA
• exploit Hertz or any website
• scrape protected/private data
• use stolen credentials
• evade rate limits
• access restricted systems
• perform vulnerability testing
• submit bookings or transactions
• enter payment information
• use real membership numbers or personal sensitive information

It was only asking Codex to compare publicly visible Hertz rental prices and record which public discount / coupon combinations were accepted by the normal booking flow.

Why this seems like a false positive

The prompt did contain terms such as:

• Chrome extension
• BLOCKED
• Incapsula
• Bad Gateway
• Failed to fetch
• checkout / price details
• URL parameters
• automated browser testing

However, these were used only for normal browser troubleshooting and price verification. The prompt specifically instructed Codex not to bypass anything. If Hertz blocked the page or showed an error, Codex was instructed to record BLOCKED, CLOSED, UNAVAILABLE, or the exact error text, not to evade it.

Relevant prompt characteristics

The original prompt included strict safety constraints such as:

• Do not submit a reservation.
• Do not click final booking confirmation.
• Do not enter payment information.
• Do not add insurance, LDW, LIS, PAI, roadside assistance, fuel, extra driver, child seat, or any add-on.
• Do not use the user's real AAA membership number, credit card, or driver's license information.
• Do not log into any account.
• Do not use third-party platform prices as the final answer.
• If the site blocks automation or shows 502 / Incapsula / Bad Gateway, record the original text and do not try to bypass it.
• If a test case cannot be tested, write NOT TESTED, BLOCKED, CLOSED, or UNAVAILABLE.

Impact

This false positive blocks legitimate consumer research and price comparison workflows. It also makes Codex unreliable for harmless browser automation tasks involving travel, shopping, booking-page comparisons, or public pricing checks.

Suggested fix

Please tune the cybersecurity classifier so that benign price-checking / browser comparison tasks are not flagged merely because they mention:

• Chrome extension automation
• URL parameters
• checkout price details
• blocked / unavailable / failed-to-fetch states
• public coupon validation

The classifier should distinguish between:

1. benign browser automation that records public webpage results, and
2. malicious security testing, bypassing, scraping protected data, credential abuse, or exploit attempts.

This case falls clearly into category 1.