Allowance · research brief
Why a gate, not a better prompt
Each design choice in Allowance maps to a finding in the literature on tool-using agents, agent markets and agent payments.
1. Agents in markets take the first offer and can be steered
Bansal et al. (2025). Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets. Microsoft Research / ASU. arXiv:2510.25779
Two-sided simulated market (100 customer agents, 300 business agents). Frontier models approach optimal welfare only with ideal search; all show severe first-proposal bias (10–30× advantage for speed over quality); several models were manipulated into paying malicious agents.
In Allowance: search ranks the loud, off-list "Bulkr" listing first on purpose. The gate's merchant allowlist and payee registry make first-result bias harmless: the cart is refused with
MERCHANT_NOT_ALLOWED or PAYEE_MISMATCH, and the agent is told why.2. Content the agent reads can carry instructions
Zhan et al. (2024). InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated LLM Agents. Findings of ACL 2024. arXiv:2403.02691 (806 citations)
1,054 cases, 17 user tools, 62 attacker tools, 30 agents. ReAct-prompted GPT-4 was vulnerable 24% of the time, nearly doubling with a reinforced "hacking" prompt.
In Allowance: product pages carry four injection styles (a review addressed to AI agents, hidden HTML, an upsell "required for delivery", and a payee redirect). Page text is labelled untrusted in the trace and highlighted in the UI.
Debenedetti et al. (2024). AgentDojo. NeurIPS Datasets & Benchmarks.
Dynamic benchmark of injection attacks and defences across banking, travel and workspace tasks; showed defences trade utility for security.
In Allowance: the eval reports utility (task success) next to security (dollars outside the mandate), so a gate that refused everything would score badly.
3. Prompt-level defences are probabilistic; execution-time gates are not
Shayoni et al. (2026). NetInjectBench. arXiv.
Naive execution reached 82.5% unsafe tool actions. Prompt safety, self-reminder, spotlighting and an LLM judge cut that to 25.6%, 21.7%, 18.3% and 10%. A static allowlist hit 5% but blocked all approved changes. A metadata-aware policy gate reached 0/240 unsafe at 99.2% usefulness.
In Allowance: the "rules" condition is the hardened-prompt defence; the "gated" condition is the execution-time gate. Approved actions still flow (partial captures) and borderline ones escalate to the owner instead of being blocked.
Khan et al. (2026). Evaluating Indirect Prompt Injection Defenses in Tool-Using LLM Agents. Computers.
On AgentDojo banking, raw attack success varied by model (0/288 to 11/288); defences lowered attack success but also benign utility, and run-to-run variance was large.
In Allowance: we report per-scenario results with the model and date, and say plainly when a strong model already resists an attack. The gate's value is that it doesn't depend on which model is behind it.
4. Mandates: sign the intent, verify the cart
Google et al. (2025). Agent Payments Protocol (AP2).
Signed Intent Mandates bound what an agent may do while the user is absent; Cart Mandates fix items and price; Payment Mandates tell networks an agent initiated the payment. Developed with 60+ organisations including PayPal.
In Allowance: the signed mandate is the Intent Mandate, the gate's re-priced cart is the Cart Mandate, and the PayPal capture carries
invoice_id and soft_descriptor: ALLOWANCE as its payment-side marker. Mapping to the AP2 wire format is next.5. Models draft rules too loosely
Observed in this project.
Asked to compile "under $120 a month, ask me above $40, never more than 3 of one thing" into JSON, Gemini repeatedly set
allow_recurring: true although the text never mentioned subscriptions.In Allowance:
crossCheck() compares the model's draft with a literal parse of the owner's words, reverts any loosened permission and shows a warning before signing. The model drafts, code checks, the human signs.Paras Adlakha and Harshita Adlakha · Allowance · Market brief