cdp

No direct malicious code (network exfiltration, credential theft, persistence) is evident in this module alone. However, it provides high-risk, manifest-driven plugin execution: it dynamically imports node-side tools and executes browser-side tool source via CDP Runtime.evaluate using code embedded from on-disk files. If an attacker can tamper with skills/cdp/learnings/ or influence domain/tool/manifest inputs, this becomes a practical supply-chain automation RCE risk. Directory/path handling for the caller-provided domain should be treated as requiring strict validation or allowlisting.

This module is security-critical: it starts a localhost HTTP server that accepts POST /eval and executes attacker-supplied JavaScript using eval() with no authentication and no sandboxing. Because it also exposes privileged session/CDP/recording and automation helpers on globalThis, an eval payload can control the connected browser/session and potentially trigger recording behaviors. A unauthenticated POST /quit can also terminate the service. Treat as a high-risk REPL/backdoor pattern; restrict access strongly or remove eval in production/supply-chain deployments.

This fragment is best characterized as an engineered media extraction/recording pipeline using CDP to pull in-memory MSE/SourceBuffer output from a target page, base64-encode it in the page, decode it in Node, and write reconstructed bytes to local disk during playback. No classic malware behavior (credential theft, command execution, persistence, or network exfiltration) is shown, but the code’s core capability enables unauthorized content duplication and privacy compromise, with additional availability risk managed by explicit slice sizing. Treat as a high misuse/privacy-risk component rather than confirmed system-compromising malware.

This code fragment is a high-risk browser instrumentation recorder. It injects document-level listeners into every CDP-attached tab (including tabs opened later), captures user clicks and keystrokes (keydown keys and modifiers) plus element/context metadata and (for non-password fields) input values and displayed text, and funnels the results into a persistent in-memory log that is later returned to the caller. There is partial masking only for password-type fields; truncation does not neutralize sensitivity. No obfuscation is present, but the behavior strongly resembles keylogging/behavioral surveillance capabilities, making it unsuitable for untrusted supply-chain inclusion without explicit consent, tight scoping, and strong safeguards.

This fragment implements a screenshot- and vision-assisted, coordinate-based CAPTCHA interaction/bypass workflow for multiple providers by locating CAPTCHA iframes, injecting synthetic mouse/text inputs to solve challenges, and iterating through dynamic state transitions. It does not exhibit traditional malware characteristics in this snippet, but the intended capability is security-abusive (bypassing anti-bot protections) and should be treated as high-risk for operational abuse.

This code fragment is a high-abuse, access-control-evasion scraping/replay technique. It captures authenticated API requests (including sensitive headers/POST bodies), then replays or rewrites them from inside the page context using Runtime.evaluate and fetch with credentials:'include', returning full responses to the caller. It is not indicative of traditional malware/persistence, but it materially enables unauthorized data extraction and bypass of UI/paywall/bot/rate-limit protections when misused.