AI Agent Uncovers Unauthenticated Router Config Dump Leading to CVE Filing
An experienced security engineer running a home lab with Proxmox VE asked an LLM agent to document router administration methods. The agent was given only basic port information and instructed to verify available functionality. Overnight the agent located an unauthenticated configuration read endpoint on the router’s web interface.
The endpoint returned approximately 37 KB of configuration data containing the administrative password in base64, all WPA2 pre-shared keys, service credentials, and full network settings. The agent confirmed that the corresponding write endpoint correctly required authentication, indicating a design flaw rather than an intentional backdoor. It then used the extracted credentials to obtain a valid session cookie, proving full administrative access was possible from any device on the LAN.
Technical details and classification
The router runs an ancient Boa web server exposing a single-page application built with jQuery. JavaScript bundles revealed the real API endpoints. The flaw was classified as CWE-306 Missing Authentication for Critical Function, with secondary weaknesses CWE-522 and CWE-359. The agent calculated a CVSS 3.1 score of 8.8 (AV:A/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H).
External scans confirmed the issue is strictly LAN-facing because WAN management remains disabled. The researcher could not locate current firmware updates due to unresponsive vendor infrastructure, a fact also included in the disclosure.
Path to CVE registration
The agent prepared a bilingual coordinated disclosure report for the vendor and assembled the required fields for the MITRE CNA-LR submission using knowledge of the Vulnogram form. The researcher personally sent the email and submitted the CVE request. A parallel entry was created in VulDB. The 90-day disclosure window is now open.
Key takeaways for agent use in security research
- Read-only permission boundaries successfully limited the agent to reconnaissance and verification steps.
- The agent performed exhaustive enumeration of JavaScript files, endpoints, and public vulnerability databases far faster than a human researcher.
- Final actions requiring accounts or submissions remained under human control.
- Home labs provide a safe, legal environment for testing autonomous security agents.
Related articles
Managing Bug Bounty Programs: Handling Report Overload, Duplicates, and AI-Generated Submissions
Launching a bug bounty program is only the beginning; the real challenge lies in managing the influx of reports, filtering duplicates and AI-generated noise, and ensuring vulnerabilities are actually fixed. Standoff Bug Bounty recommends starting with private programs limited to 30-50 researchers to build internal processes before scaling. Companies like Flavau developed custom DevSecOps consoles that cut report triage time by a factor of 10 by automating duplicate detection and assignee assignment. The platform introduced neural slop windows that limit submissions based on researcher quality history, addressing the surge in AI-generated invalid reports observed in spring 2026. Fixed response timelines, clear ownership roles, and prioritized remediation of critical issues help maintain program effectiveness. Google’s decision to pause its OSS VRP in October 2026 due to automated report floods highlights broader industry concerns, yet experts maintain that human-driven independent validation remains essential. Key success metrics include reduced risk from fixed high-severity issues, lower average payout costs, and improved cross-team communication between security and development.
Researcher Bypasses Safe Exam Browser Checks in Moodle Using MITM Proxy and Canonical Config Key Hash
A detailed technical write-up describes how a student bypassed the proprietary Safe Exam Browser requirement for Moodle quizzes by reverse-engineering the single-header authentication mechanism. The bypass relies on fetching the public SEB configuration XML from the open endpoint, computing a canonical SHA256 hash known as ConfigKey, and injecting the resulting X-SafeExamBrowser-ConfigKeyHash header via mitmproxy. Moodle caches a successful check inside the PHP session, allowing all subsequent requests to be made from a normal browser without further header manipulation. Two implementations were created: a headless Python client for full automation of quizzes and a transparent mitmproxy addon that lets a user continue interacting with the test manually. The author also handled edge cases such as crossword OCR with Tesseract, dynamic qubaid and sequencecheck values on retakes, and audio listening sections. The resulting scores reached 772 out of 780 points, placing the account 22nd out of 372 students. The full code, including configkey.py, mitm_addon.py and unit tests, has been published under CC BY-NC-SA 4.0.
Oracle Releases 655 Security Patches in November 2026 Critical Update
Oracle has published its November 2026 Critical Security Patch Update, delivering a total of 655 patches that address vulnerabilities across a wide range of products. The update resolves 640 unique CVEs after accounting for duplicates and supplements the regular quarterly Critical Patch Update. Affected products include Oracle Database Server, Oracle Java SE, MySQL Server, Oracle WebLogic Server, Oracle E-Business Suite, and Oracle VM VirtualBox. Among the fixes, Oracle Hyperion accounts for the largest share with 267 patches, followed by Oracle Supply Chain with 74 and both Oracle Fusion Middleware and Oracle E-Business Suite with 71 each. Of the vulnerabilities patched, 311 carry a CVSSv3.1 base score of 7.0 or higher, including 62 rated 9.0 or above, while 303 can be exploited remotely.
DPoP Enforcement Failure Allows Bearer Token Bypass in redb.Identity OpenID Server
A vulnerability in redb.Identity allowed DPoP-bound access tokens to be accepted without proof-of-possession on the /connect/userinfo and management API endpoints. The flaw affected versions 1.0.1 through 4.6.1 when DPoP was enabled and was assigned GHSA-vmj2-8phj-mfxf with a CVSS 3.1 score of 7.4. External researcher Rathnadeepan reported the issue privately on 4 November, enabling a same-day confirmation and a fix released in 4.6.2 just one day later. Existing test suites, including over 2,000 green tests across three databases, failed to catch the problem because they only verified positive cases and never tested rejection of missing or invalid DPoP proofs. The root cause traced back to endpoints that only inspected the Bearer scheme and an earlier OpenIddict migration change that disabled cnf validation for jkt-bound tokens. The project published the advisory without waiting for a CVE and added negative test cases plus demo scenarios to prevent similar gaps in the future.