Commenter proposes redesigned CAPTCHAs and age-verification links to counter vision-language models that solve CAPTCHAs
A reader comment by Clive Robinson on Schneier on Security discusses whether legally mandated age-verification checks and corporate surveillance systems could make CAPTCHAs obsolete as AI models learn to solve them. He suggests that age-verification documents, limited to natural persons, may not be available to current AI systems, and he warns that any leak of such documents to AI operators would defeat the purpose of the checks. He also proposes a CAPTCHA redesign. This is opinion and speculation, not a reported incident, and it has no direct threat impact.
Technical Details
- Proposed design: a larger 5x5 image grid with three image types: (1) contains the requested object, (2) does not contain it, (3) ambiguous.
- A fully trained VLM would select only Type 1 images and get all of them right. A human would be expected to make occasional errors on Type 2 and Type 3 images.
- The backend would look for a small percentage of errors on Type 2 and Type 3 images and treat too few errors as bot behaviour.
- Three or four CAPTCHA frames would be submitted in series and checked together as a group, to reduce random probing by AI systems.
- The commenter notes that chained agentic AI systems could be used to fetch age-restricted content and pass it back, bypassing age verification. He compares this to search engines being used to get around URL-based blocking in national firewalls.
Remediation
No remediation is stated. The article offers only the conceptual CAPTCHA redesign ideas above.
Linha do tempo
-
Commenter proposes redesigned CAPTCHAs and age-verification links to counter vision-language models that solve CAPTCHAs
ler a análise desta fonte ocultar
A reader comment by Clive Robinson on Schneier on Security discusses whether legally mandated age-verification checks and corporate surveillance systems could make CAPTCHAs obsolete as AI models learn to solve them. He suggests that age-verification documents, limited to natural persons, may not be available to current AI systems, and he warns that any leak of such documents to AI operators would defeat the purpose of the checks. He also proposes a CAPTCHA redesign. This is opinion and speculation, not a reported incident, and it has no direct threat impact.
Technical Details
- Proposed design: a larger 5x5 image grid with three image types: (1) contains the requested object, (2) does not contain it, (3) ambiguous.
- A fully trained VLM would select only Type 1 images and get all of them right. A human would be expected to make occasional errors on Type 2 and Type 3 images.
- The backend would look for a small percentage of errors on Type 2 and Type 3 images and treat too few errors as bot behaviour.
- Three or four CAPTCHA frames would be submitted in series and checked together as a group, to reduce random probing by AI systems.
- The commenter notes that chained agentic AI systems could be used to fetch age-restricted content and pass it back, bypassing age verification. He compares this to search engines being used to get around URL-based blocking in national firewalls.
Remediation
No remediation is stated. The article offers only the conceptual CAPTCHA redesign ideas above.