MAQPNA

Field notes · · 4 min read

Know Your Agent verification raises AI agent security stakes

Only a few hundred AI‑driven orders have reached DoorDash, yet a single verified agent could open a bank account and transfer funds on behalf of a CEO.

Verification is no longer optional

Shufti’s newly announced Know Your Agent (KYA) service makes the first serious attempt to tie an autonomous AI agent to a verified owner and a set of explicit permissions. The product recognises that an agent that can open accounts, place orders or move funds must be provably linked to a real person or business, and that the receiving party needs to know who is accountable before it acts on a request.

KYA does not rely on a simple access token. It adds layers of identity proof – similar to the way banks verify a human customer – and then issues a short‑lived token that encodes both the owner’s identity and the actions they have authorised. In practice this means the downstream service can check, “Is this request coming from a KYA‑approved agent, and does the owner’s policy allow this exact operation?”

Agents are already touching real services

DoorDash confirmed that it has already received “several hundred” orders generated by AI agents. The company is open to third‑party agent makers, but the volume shows that production agents are moving beyond test labs into consumer‑facing commerce.

Chainalysis disclosed an intrusion in which an external actor leveraged an “offensive agentic” tool to access internal systems and scrape non‑sensitive customer identifiers. Although the breach did not expose authentication credentials, the incident highlights how an unauthorised agent can become a foothold for a broader attack chain.

The week in the feeds · 7 days to 2026-11-01184 stories across 10 sources, by topic
Vulnerabilities44Data leaks26AI agents16Supply chain8Regulation & governance3Prompt injection1Sovereignty0

A story can count under more than one topic. Counted from the feeds, not estimated: Dark Reading, Hacker News, Help Net Security, Infosecurity Magazine, SecurityWeek, Simon Willison, TechCrunch, The Hacker News, The Record, The Register.

The security gap: agents can act without policy enforcement

Google’s Gemini 4 Carbon prototype demonstrated dramatic gains in agentic tasks, including vulnerability discovery and exploit generation. The same capabilities that help defenders can empower malicious agents to probe, manipulate and exfiltrate data at speed.

When an agent can call tools, invoke models or execute code without a gate, the organisation loses the ability to audit intent, enforce least‑privilege, or stop a dangerous operation before it happens. Traditional IAM controls do not understand the nuance of “agent‑to‑tool” calls, so they treat an agent like any other service account – a risky shortcut.

Three concrete controls to close the gap today

Start with a default‑deny policy at the gateway that sits in front of every tool, model and external API. Only calls that match an explicit rule are allowed, and any mismatch is logged and blocked.

Introduce a four‑eyes approval workflow for high‑value operations such as payments, account creation or data export. The approval should be tied to a separate identity from the agent, ensuring segregation of duties.

Deploy a kill‑switch that can revoke all active agents in milliseconds.

  • Pin MCP tool definitions to avoid silent “rug‑pull” changes.
  • Enable prompt‑injection guards (Prompt Guard, Lakera, Bedrock) on every inbound prompt.
  • Enforce hierarchical budgets and rate limits to cap spend and data egress.

Where MAQPNA fits into a verified‑agent strategy

MAQPNA is the sovereign runtime for AI agents: it checks, approves and records every action AI agents take, on infrastructure the customer controls. By placing a single gateway between agents and every tool, model or other agent, it lets you enforce the default‑deny policies, four‑eyes approvals and kill‑switch described above while producing a SHA‑256 hash‑chained audit ledger that survives any attempt to tamper with the record.

The paceStories per day, last 7 days
13Mon22Tue27Wed33Thu68Fri20Sat0Sun

Takeaways

  • Adopt a KYA‑style identity token that binds each agent to a verified owner and explicit permissions.
  • Place a policy gateway that defaults to deny and requires a rule for every tool or model call.
  • Require two‑person approval for any request that creates accounts, moves money or extracts data.
  • Implement a rapid kill‑switch and test it – MAQPNA demonstrates a 302 ms response at p99.
  • Log every decision to a tamper‑evident ledger; use the ledger to satisfy audit requirements such as the EU AI Act.

Questions people ask

How can I prove an AI agent is acting on behalf of a real user?

Use a verification service such as Shufti’s Know Your Agent to issue short‑lived identity tokens that embed the owner’s credentials and allowed actions. The downstream service checks the token before executing any request.

What is the fastest way to stop a rogue agent in production?

Deploy a kill‑switch at the MAQPNA gateway.

Sources

  1. Shufti ties AI agent actions to verified owners and permissions · Help Net Security
  2. DoorDash said it’s only received ‘several hundred’ orders from AI agents · TechCrunch
  3. Chainalysis Confirms Intrusion Into Internal Tools and Exposure of Customer Identity Data · The Hacker News
  4. Google's Gemini 4 Carbon Leads Benchmarks for Finding, Exploiting, and Patching Vulnerabilities · The Hacker News

Written with Cloudflare Workers AI from the stories above, then checked automatically: every source is a real link from that day's feeds, every number is in a source or our fact sheet, and nothing is copied. Spotted a mistake? Tell us.

Let agents act. Keep control.

Run MAQPNA on a laptop in minutes, then in your own cloud, cluster or data centre.

More install options
curl -fsSL https://maqpna.com/install.sh | sh