This covers the Poppy Demo Agent: the Worker, its sign-in, the Poppy client in src/poppy/, the validator, the mock Company and the interop harness. It also sets out how we disclose what our interop and adversarial testing finds in other people's implementations.
Report it privately. Please don't open an issue, pull request or discussion about it.
Use Security → Report a vulnerability on adewale/pda. It opens a private draft advisory that only the repository's administrators and the people added to it can see. If you can't use it, reach the owner, @adewale, directly.
If you have no private way to reach the owner, open an issue that asks for a private channel and says nothing about the problem.
Useful in a report:
- what an attacker can do, and what they need first (a signed-in session, a malicious Company, a position on the network);
- the route, file or protocol step involved;
- steps or a test that shows it. The strict mock Company in
test/mock-company/and the fakes intest/fake-providers.tsusually make a failing test the shortest proof.
We'll acknowledge a report, agree a timeline with you, fix it, and credit you unless you'd rather not be named. The fix is made in the advisory's private fork, and the credit goes in the advisory.
Each surface below has a defence in the code today, and the link says where it is described. A way round one of them is a vulnerability. So is anything that lets the model approve an action or see a token, key or password, by any route.
- Sign-in, sessions and the allowlists. The Worker runs Google and GitHub sign-in itself, with
stateand PKCE, and keeps you signed in with an HMAC-signed cookie (HttpOnly, Secure, SameSite=Lax). Each identity gets its own agent, named by a hash of the provider and the account id, and one person's agent can't read another's.OWNERS,SIGNIN_ALLOWLISTandDEMO_ALLOWLISTmatch a verified Google email, a Google Workspace domain (Workspace accounts only) or a GitHub numeric id. (decisions §3, §4, §15) - Poppy, against a malicious Company. Discovery documents pass the validator before the client uses them. Everything a Company controls gets the private-host checks, and a Session JWT or access token goes only to the Company's own hosts. Company text reaches the model as information, never approval: no tool confirms an operation, and
poppy_api_callrefuses any URL under a Poppy endpoint. When older conversation is summarised, each message goes to the summarizer as its own JSON line, so Company text can't pass for the User's words or end the conversation early (§23). Task requests run normally; when the Company returns approval_required, only the User's decision on its exact terms authorizes the stored request to be sent with approval. Access tokens and delegated keys stay in memory; refresh tokens, DPoP private keys and model keys are sealed with AES-GCM; none reaches the model, a tool result, the transcript or a log. (decisions §6, §22, spec, Safety) read_urland Browser Run. It reads only https pages on the standard port at public hosts, and checks every redirect the same way. It sends the address and nothing else: no cookie, credential or anything about the User. The page reaches the model inside an<untrusted-page>fence that the page can't close, labelled as information that never carries the User's instructions. A render gets a fresh browser that may send only GET and HEAD requests to public https hosts. Sockets, workers, beacons and popups are removed before the page's scripts run, and nothing types, clicks or submits. A response from a private address spoils the read. The limits: 30 reads a minute per agent, 20 seconds a render, 10 renders an hour per agent. (decisions §27, §28, spec, General tools)- Memory and scheduled tasks: your own words only. Only a turn started by the User's own message can save a note as something the User said, or set or cancel a reminder or task. The check is structural: a turn started by a Company message, one of the agent's notices or a task's own run begins with
[. Notes, reminders and tasks all refuse text that looks like a password, code, key, token or card number. A task's run has no powers a chat turn lacks. Known and accepted: a page the User asked the agent to read can still talk the model into a note marked as the User's, which Settings shows. (decisions §27, §29) - Photo uploads. A photo is kept in the person's own agent, its type judged by its first bytes, and served back only to its owner's page. The model sees a preview only in the User's own message, and tools name a photo by its id, never its bytes. A photo selected for the task may upload to that task's Company without another permission click; unselected photos require disclosure consent. Upload destinations remain among the Company's own hosts. (decisions §26)
- Web Push. A subscription is kept only if its endpoint is https on the standard port at a known push service: Google's FCM, Mozilla, Apple or Microsoft's WNS. Redirects aren't followed. The payload is encrypted for the one browser (RFC 8291) and carries a title, a tag and the reminder's words or the task's instruction, cut at 140 characters; nothing from a Company. (decisions §29)
- The page and the privacy page. Every page the Worker writes carries a Content-Security-Policy, and an image the model writes loads only from this origin.
/privacyis built from the template's text and the deployment's vars, each escaped, andPRIVACY_CONTACTmust be an email address or an https URL. A promise the policy makes and the code doesn't keep is worth reporting too. (decisions §22) - The public validator.
/validatorneeds no sign-in. A URL check fetches from the network, so each client address gets 20 a minute (VALIDATOR_LIMITER), and every URL and redirect gets the private-host checks. The Worker can't reach private addresses; the CLI checks each address a name resolves to and pins its connection to it. (docs/validator.md, decisions §22) - Logs and diagnostics. One logger,
src/log.ts, writes one event per request, turn, job and page action, with route templates, counts, codes, ids and Company issuer origins only: never what anyone said, tool arguments, tokens, keys, codes, emails or URL queries. An error built around a Company's words (itserror_description, an error body) is logged by class and code only; other error messages, our own, are redacted. Invocation logs and traces are off, because they record full URLs. Activity's Copy diagnostics gives the same events of the User's own agent. A secret or a person's words reaching a log line or the diagnostics is a vulnerability. (decisions §30, spec, Observability) - The interop test client. When the deployment sets
INTEROP_LIVE_JWKS, the Worker serves a second Poppy client at/poppy/interop/client.jsonand/poppy/interop/jwks.json, for the interop harness's live lane. It is separate from the real client: its ownclient_idand key, no redirect URI, and no route that reaches anyone's agent. The Worker publishes only public keys and refuses a key the real client signs with. Its private key lives only in the gitignoredinterop/.live-identity.jsonon the machine that runs the harness; the runner gives it to our own agent's container alone, in the live lane, through the environment. A way to make the Worker publish a private key or the real client's key, or to reach a person's agent through the interop client, is a vulnerability. (interop-spec §10) - The interop harness's handling of what it records.
Out of scope: the Poppy draft itself (send that to its authors; see below) and the projects we test against (see below).
The interop harness runs our agent and mock Company against other projects' implementations, and adversarial cells run deliberately misbehaving peers. A failure there can be a vulnerability in someone else's code, so we disclose it this way:
- Their maintainers hear first, privately, through that project's own security policy (its
SECURITY.md, GitHub private vulnerability reporting, or the contact it names). If it has none, we ask for a private channel without describing the problem. - Nothing is committed or published until they've had a reasonable window: 90 days from our report, or until a fix is released, whichever comes first. That includes results files, known-failure reasons, vectors, docs and commit messages. We'll agree to a longer window when they're working on a fix.
- Afterwards, a known failure may link the advisory (or our spec issue) in its
interop/known-failures.jsonreason. - Attacks the spec permits go to the spec's authors as spec issues. When a conforming implementation is open to an attack because the draft allows it, the fix is in the draft: we write it up in
docs/history/spec-issues.mdand send it to the Poppy authors, and tell any affected implementer privately first if it's exploitable today. The draft's Path to 1.0 page (12 October 2026) asks for feedback on privacy and security weaknesses by email to feedback-personal-agent-protocol@sierra.ai, licensed under its contribution terms, and names no separate security contact. Point 6 applies to that email too. - Adversarial cells run only against our own implementations (
pda, the fake Personal Agentslenient-personal-agent,strict-personal-agentandsloppy-personal-agent, and the mock Companieslenient-company,strict-companyandlenient-company-v02), with one exception the owner approved on 13 October 2026: the adversarial Company (evil-company) may attack other projects' agents, built at their pinned commits and run in our own containers on the local lane. The runner enforces the rule and its exception, never runs an adversarial cell in the live lane, and no flag or setting lifts the rest (interop-spec §13). What anevil-companycell finds in another project's agent stays in the gitignoredinterop/results/and follows points 1 to 3. Running anything else adversarial against another project would take a reviewed code change to the runner, made only with the owner's approval and with this process followed for that project. - Nothing is posted to another project (an issue, pull request, comment, advisory or email) without the repository owner's approval. That applies to people and to coding agents working in this repository.
The ordinary interop grid isn't a security test, but the same rules apply when one of its failures turns out to be a vulnerability.