Amine Affif
A full-stack developer based in Paris who focuses on security architecture for AI agent systems. He advocates placing security boundaries at the process startup stage rather than within prompts, an engineering practice aimed at hardening agent deployments.
Timeline (last 90 days)
Amine Affif is a full-stack developer in Paris.
The Agent built by Amine Affif uses a closed tool list: 12 read tools are allowed in advance, and 15 write tools are denied.
Due to the tool list restriction, Affif's Agent cannot write files or send messages.
In addition to the hard tool-list switch, Affif layered on three protective measures: an anti-injection prompt section, mandatory adversarial checks, and human verification.
Paris-based full-stack developer Amine Affif argues in his blog that security restrictions on AI Agent tools must be set at the process startup stage, not stuffed into the prompt.
Affif advises: before connecting any tool to an Agent, first figure out clearly what capabilities it should not have, then ensure at the architectural level that these capabilities do not exist.
Affif shares the actual configuration of the Agent he built: the tool list was fully locked before startup, 12 read tools were authorized, and 15 write tools were denied.
On top of the tool allowlist, Affif layers three protective measures: an anti-injection zone in the prompt, adversarial review before output, and human confirmation before release.
Affif states explicitly that he cares less about the efficiency gains brought by tools than about how to effectively contain these tools.
All Facts (9)
Amine Affif is a full-stack developer in Paris.
50%UnverifiedThe Agent built by Amine Affif uses a closed tool list: 12 read tools are allowed in advance, and 15 write tools are denied.
50%UnverifiedDue to the tool list restriction, Affif's Agent cannot write files or send messages.
50%UnverifiedIn addition to the hard tool-list switch, Affif layered on three protective measures: an anti-injection prompt section, mandatory adversarial checks, and human verification.
50%UnverifiedParis-based full-stack developer Amine Affif argues in his blog that security restrictions on AI Agent tools must be set at the process startup stage, not stuffed into the prompt.
50%UnverifiedAffif advises: before connecting any tool to an Agent, first figure out clearly what capabilities it should not have, then ensure at the architectural level that these capabilities do not exist.
50%UnverifiedAffif shares the actual configuration of the Agent he built: the tool list was fully locked before startup, 12 read tools were authorized, and 15 write tools were denied.
50%UnverifiedOn top of the tool allowlist, Affif layers three protective measures: an anti-injection zone in the prompt, adversarial review before output, and human confirmation before release.
50%UnverifiedAffif states explicitly that he cares less about the efficiency gains brought by tools than about how to effectively contain these tools.
50%