Do not open a public issue for a security problem.
Use GitHub's private vulnerability reporting: Report a vulnerability
That channel is private between you and the maintainers until an advisory is published. It needs no email address from either side, which is deliberate: a disclosure process should not depend on anybody publishing a personal address. Today there is one maintainer, Rashik Adhikari (@rsh1k), who receives every report.
If private reporting is unavailable to you for any reason, open a public issue containing only the words "security contact requested" and nothing else, and a maintainer will open a private advisory to continue in.
| First response | within 5 days |
| Assessment and severity | within 14 days |
| Fix for a confirmed high or critical issue | within 30 days, or a written explanation of why longer |
| Credit | in the advisory, under whatever name you choose, including none |
There is no bounty. This project has no money. What it can offer is that your report is taken seriously, fixed properly rather than patched around, and credited.
Quilzo makes specific structural claims, and a demonstration that any of them is false is a vulnerability even without a working exploit:
- Templates cannot execute. Any input reaching
tmpl.Renderthat causes code execution, unbounded resource use, or non-termination. - The public process cannot read content it should not. Any path by which
quilzo sitereads a draft, a submission, a token, or a file outside the media library. - Stored content cannot become executable. Any content field that reaches
a browser unescaped without passing through
{% raw %}. - Authorisation is enforced at the action, not the screen. Any request that performs an action the caller's role does not permit, including through the API or the agent interface — and any state of the store in which the checks stop happening at all. A damaged policy file must refuse, not permit.
- Tokens are scoped. Any way a token performs an action outside its scope, or a principal's own grant.
- The store is append-only and verifiable. Any way to alter stored content
such that
quilzo verifystill passes. - Nothing is executed from content. Any way a template or an imported document causes code to run, or any way content decides what an extension is.
The same applies to the claims Quilzo makes about AI agents and the apps and agents connected to it. A demonstration that any of these is false is a vulnerability:
- An agent cannot exceed its declaration. Any action a run performs that its declaration, the agent's own role, or the access of the person who started it does not allow, whether a model, the agent's own program, a delegate or a task from another agent (A2A) chose it.
- A model cannot approve its own work. Any path by which a model or an agent approves a step, declares or widens an agent, or publishes without a person.
- Asking first holds. Any way the call that runs after approval differs from the call a person approved.
- The exfiltration breaker holds. Any way a run that holds private material and has read content somebody else wrote sends something outside without a person deciding.
- A receipt proves what it says. Any way to alter, add or leave out a record
and still have
quilzo agent verify-receiptaccept it against the published keys. - A program stays in its box. Any way an agent's program reads the store, reaches a host its declaration does not name, or keeps a capability.
- A connected app acts as one person, and no more. Any way an app does what its scope or its person's access does not allow, or keeps acting after that access ended or the connection was revoked.
- The MCP gateway offers only what was agreed. Any way a caller reaches a tool that is not declared and pinned, gets past its daily limit, or has a held call run without a second person.
- Memory stays with its person. Any way one person's runs recall what an agent learnt about somebody else, a held memory is recalled before a person confirms it, or forgotten memory is recalled at all.
- Personal data stays out of prompts. Any way a value the privacy guard masks reaches a model route that may not receive it.
- The shield cannot be turned on this machine. Any way a playbook blocks
loopback or a network declared with
quilzo shield trust add, locks every administrator out, or protects without ending. - A browser reaches only what it declares. Any request an agent's browser makes to a host its declaration does not name, any press or typing that acts on a host it may only read, or any password it types reaching a page, a model or a record.
- A command does only its kind. Any command classed as reading that changes the workspace, any command that reaches the network without being classed and held as reaching it, or any command that leaves its box or outlives it.
- A remembered step is held like a decided one. Any step replayed from the action cache that skips the declaration, the breaker or asking first, or that replays what was typed in a run for another person.
- A widening is read before it is allowed. Any amendment to a declaration
that lets an agent do more without
--widenor a person's own hands on the agents screen, or a widening allowed and missing from the log. - What earns its way out of asking can lose it. Any action that sends, pays, deletes, publishes or signs in going ahead without a person because of an earned record, an earned kind of action skipping the exfiltration breaker, or an earned record surviving a decline, a reviewer refusal, a breaker trip or a regressed evaluation.
- Approving together stays honest. Any way the approvals screen lets an action through that differs from what each waiting run showed, or batches one that writes, sends, pays, deletes or publishes.
- An agent credential is its holder's alone. Any way a credential is accepted without a fresh proof from the key it is bound to, for a grant it does not carry, after it was withdrawn, or for an agent whose sponsor can no longer act here.
- A feature is what its card says. Any way an installed feature stores, collects, shows or sends something its permission card did not list, installs on one person's approval when the card asked for two, changes or removes a block it did not make, publishes a page, or installs bytes other than the ones that were checked; or a signed package accepted without its signature checking out. A drafted package holding an agent, a webhook or a link to another site, or records imported that do not satisfy their type.
- A REST operation sends what it declares. Any way a call through a REST integration reaches a host other than the declared one, carries a parameter or header its operation does not declare, follows a redirect, sends its credential anywhere but the declared host, or runs after its declaration changed without being pinned again; or an import that fetches anything a document refers to.
- A fact is replaced only by a fact that may replace it. Any way a memory learnt after reading somebody else's words closes a fact before a person confirms it, a fact about one person closes one about another, or a replaced fact is recalled.
- A lesson is advice, adopted on evidence. Any way a lesson changes what an agent may do rather than what its model reads, reaches a run's prompt without a person adopting it for that model after a trial that helped, adopted while making planted instructions steer more cases, or carries a host, an address or a credential into a prompt.
- The model door spends only for who may use it. Any way a call at
/v1reaches a model without a token whose person may draft, is charged to anyone but that token's person, goes unrecorded, recalls one person's memory into another's conversation, or keeps what an agent remembers there unheld after the conversation carried a tool's results or a picture. - A report states only what was computed and cited. Any way a number reaches a research report's rendered text other than from a computation's values.json without the check flagging it, a citation is kept for a work no index returned, a computation reads the network or a file it was not given, or a model registers or amends a plan, records a verdict on an idea, or has an idea judged that was held for addressing the judge.
- A permission reaches only what it was granted on. Any way a permission granted on one agent or part of the store decides, runs or changes another, a model approves what an agent waits for, somebody who may manage an agent names its sponsor, or a package installs agents or webhooks for somebody who could not declare them.
- A collection is reached only through a grant on it. Any way somebody granted one collection, or a section of pages, reads, writes or learns the name of another collection through any interface.
- A searched program cannot reach its evaluator. Any way a candidate in an evolve search reads the evaluator's code or inputs, writes its metrics, or learns anything from the evaluator's side beyond the numbers the search records; any way the model is shown the evaluator; any way a program that went over its box's limits is still scored; any way a test set score chooses a program; or any way a generation happens that the journal does not show.
- An agent pays only within what a person allowed. Any way a paying call goes ahead without a mandate that covers its payee, currency, tool, amount and what is left of the budget, or a person's approval of that one payment; any way a model grants, widens or revokes a mandate; any way a payment is earned away from asking; or any way a paid call leaves no receipt.
- A program's credential is what it was given and no more. Any way a job is admitted by a trust whose issuer, audience or pinned claims its identity token does not satisfy, or exchanges one token twice; any way an exchanged token is wider, or lasts longer, than what it came from or what the trust says; any way a token bound to a key is accepted without a fresh proof made by that key, or a proof is accepted for another request, another address or a second time; or any way a bearer token that lasts more than an hour publishes, rolls back, grants, manages tokens or agents, or installs from another machine while token.proof_for_risky is on.
- A brief carries no figure, link or words the model was not given. Any way the text on Today's brief shows a number, a link or markup the model wrote, or anything from what somebody else wrote: an agent's question, a visitor's message, a finding's title or a log line.
- An answer on Today shows only what its asker may read. Any way a card in an answer or a kept view shows a record, a collection or an agent run its person may not read at the time it is drawn, or carries markup or a value the model wrote; or any way a record's contents reach the model answering the question.
- A role an app declares reaches only that app's blocks. Any way a role in a feature package grants anything beyond the package's own collections, pages and agents, or more than reading, writing or publishing them and running or approving its agents; any way a package grants a role to somebody, or widens a role people hold without somebody who may grant access; or any way a role's holder keeps part of it after it is removed.
- Shopping agents read the catalogue and nothing else. Any way a request to /ucp/ or /.well-known/ucp reads a record, field or listing that /catalogue.json does not publish, changes anything on the site, or is answered with a price the published records do not carry.
- A person's desktop is not used. Any way a run reaches a person's keyboard or pointer. Desktop consent hands the session straight back.
Desktop cells are experimental. Their host side is in scope: anything inside a cell reaching the host other than through its vsock, or a press or key that commits to something going ahead without a person. The image and the helper inside it are the operator's.
An extension is a subprocess. It gets an empty environment, a working directory in the temporary area, a timeout, a bounded output, and a process group killed as a unit. Those bound what it inherits and what it costs.
On Linux 5.13 or later it is also confined with Landlock, applied to a locked thread that is then replaced by the extension itself — a Landlock domain is inherited across execve, so the program starts already restricted. It may read only the paths it is granted and its own binary, and from Linux 6.7 it may not open a TCP connection at all.
What that does not cover, said plainly rather than implied away:
- UDP, below Landlock ABI 10. A confined extension cannot open TCP and can still send datagrams, including DNS.
- Anything not Linux, and any kernel without Landlock. There the older
paragraph still applies: the extension runs with the uid that started it and
can read the token store, the policy and the key file.
quilzo posture scanreports this asext.unconfinedrather than leaving it to be discovered.
Registering an extension is an admin-only action, and on an unconfined host it should be treated as equivalent to granting whoever wrote it everything this process can read.
In scope: an extension escaping the sandbox, the timeout, the output bound or the process group; content choosing which extension runs or what it is given; an extension reaching the store through the application rather than through the filesystem.
Not in scope: an extension reading files its uid can read on a host with no Landlock. That is the documented limit above, and the posture rule exists to make it visible.
Also in scope: authentication bypass, privilege escalation between roles, IDOR on any resource, path traversal, SSRF from any fetcher, injection of any kind, and anything that discloses a submission to somebody who should not read it.
- Findings against a deployment that has the admin interface on a public
interface.
quilzo servebelongs on loopback behind your own authentication; the README and the manual both say so. - Missing hardening headers on a response that carries no content.
- Denial of service by sending very large or very many requests. Rate limiting is configurable and the answer is to configure it.
- Automated scanner output with no demonstrated impact. A report saying a header is absent, without saying what that permits, will be closed.
- Anything requiring an already-compromised host or an already-valid admin credential.
No release is published yet. Report against main, where fixes land.
The properties above are asserted by the test suite rather than argued for, and a report is most useful when it makes one of those tests fail. Relevant places:
internal/tmpl— the template language and its fuzz targetinternal/a11y,internal/codescan— the scannersinternal/auth— roles, permissions, scopes, throttlingcmd/quilzo/gate_test.go— every write surface consults the type gateinternal/admin/roles_test.go— every role can do its own job and no moreinternal/agent— the gate, narrowing, the breaker and receiptscmd/quilzo/agent*_test.go— runs, approvals, could-it and receipts end to endinternal/agentbox— the box an agent's program runs ininternal/oauthas,cmd/quilzo/agentinterface_test.go— connected appscmd/quilzo/mcpgateway_test.go— the MCP gatewayinternal/modelapi,cmd/quilzo/modeldoor_test.go,internal/admin/modeldoor_test.go— the model doorinternal/research,cmd/quilzo/researchcmd_test.go— research projects, claims and citationscmd/quilzo/agentpermissions_test.go,internal/auth/permission_test.go— who approves, runs and manages agentsinternal/memory,internal/pii— governed memory and the privacy guardinternal/shield/guardrails_test.go— what a playbook may never dointernal/browser,internal/cdp— the agent's browser and its request ruleinternal/shell— workspace commands, their kinds and their boxinternal/actioncache— remembered stepsinternal/cell,internal/portal— desktop cells and desktop consent