2026-09-21 13:21 UTC
๐ชฐ We taught a fly to analyze phishing in #ANYRUN Sandbox.
No OCR, no LLM, no DOM hints, and no ability to read.
Thanks to sandbox interactivity, the fly can move through the attack chain on its own and observe the full phishing flow: follow links, resolve CAPTCHAs, go through redirects, and fill forms with an email and password.
See how the fly moves through an AiTM phishing flow:
๐ชฐ https://app.any.run/tasks/92c4e053-5abf-428a-8587-bc8abd43c7fd/?utm_source=mastodon&utm_medium=post&utm_campaign=fly_analyzed_phishing&utm_term=210926&utm_content=linktoservice
๐ชฐ https://app.any.run/tasks/f32ec968-3f2d-49f0-8654-6581370740f1/?utm_source=mastodon&utm_medium=post&utm_campaign=fly_analyzed_phishing&utm_term=210926&utm_content=linktoservice
Device Code Flow:
๐ชฐ https://app.any.run/tasks/3efc3564-7a8b-4475-8114-c86fde18ad48/?utm_source=mastodon&utm_medium=post&utm_campaign=fly_analyzed_phishing&utm_term=210926&utm_content=linktoservice
๐ง Whatโs inside:
1. ๐๐ฟ๐ฎ๐ถ๐ป. Fruit fly hemibrain mapped to FlyWire for 3D visualization.
2. ๐๐น๐ถ๐ด๐ต๐. There is inertia, sharp saccades instead of a smooth path, and braking based on how fast the image moves.
3. ๐ฉ๐ถ๐๐ถ๐ผ๐ป, ๐๐บ๐ฒ๐น๐น, ๐ฎ๐ป๐ฑ ๐๐ฎ๐๐๐ฒ. The fly cannot read. There is only contrast, movement, color, and shape. A button, link, and checkbox are different visual patterns converted into sensory signals.
4. ๐๐ฒ๐ด๐. Landing on the right area counts as a click. If the fly finds an empty field, it โsecretes acidโ: select all โ paste email.
5. ๐๐ผ๐ฝ๐ฎ๐บ๐ถ๐ป๐ฒ, ๐บ๐ผ๐๐ถ๐๐ฎ๐๐ถ๐ผ๐ป, ๐ฎ๐ป๐ฑ ๐ฎ๐๐ฒ๐ฟ๐๐ถ๐ผ๐ป. When fed, it mostly flies around and looks at the page instead of clicking blindly. When hungry, it is more active and gets frustrated faster when it finds nothing. A successful click gives a reward signal and reduces frustration. A dead click increases avoidance of that area. When the search reaches a dead end, frustration makes the fly scroll down.
6. ๐ ๐ฒ๐บ๐ผ๐ฟ๐. It remembers the screen map and does not return to places where nothing happenedโ inhibition of return.
โ๏ธ The fly is not smarter or more useful than an ML model trained for the task. This is not a phishing detection solution and not an attempt to replace one.
Made for fun, research curiosity, and the community ๐ค
#cybersecurity #infosec
Replies (1)
-
@anyrun_app@infosec.exchange 2026-09-21 13:21
The fly was supposed to analyze phishing but started doomscrolling instead. Very realistic user behavior ๐ชฐ