AstralMaris design series

The Airlock

How a fleet's agents read the web without the web reading them

OCT 2026 · Revision 1 · design paper

On this page
Full glossary · 7 terms
seat
One agent at work: a harness, its model, and the environment it runs in.
service principal
A fleet identity with no free-prose role: a typed request in, a typed result out. The build service is the first; the airlock is the second.
airlock
The service. Chamber — the five stages a page passes through. Quarantine — where the raw page is kept by hash and never handed to a seat.
finding
One suspicious or removed thing: a hidden element, a sentence addressed to “you, the assistant”, an image link to a stranger’s server.
pressure
How many findings, weighted. Equalize — pressure below the threshold; the artifact is admitted.
marked turn
A seat’s turn after reading an airlock artifact; watched more closely than other turns.
prompt injection
Text written to be obeyed by the agent reading it rather than by a person.

Revision r1 · October 2026 · AstralMaris design series The design as it stands when every web read by a seat goes through the airlock. Companion papers: Container Seats; The Build Service; Trust at the Door.


0. The one-paragraph version

Agents need to read the web, and the web is where instructions get planted for them to find. The airlock is a small fleet service — no model of its own, or only a captive one — that goes outside on a seat’s behalf, brings the page into a sealed chamber, strips it to plain text, examines it for hidden matter and planted instructions, and lets through only a typed, labelled artifact once the pressure (the weighted count of suspicious findings) is below the seat’s threshold. The seat’s own network never opens. And because no examination catches everything, the turn in which a seat reads the artifact is watched: any action outside its orders is a stop.

The operator’s charter, in his words: “the equivalent metaphor for an air-lock for incoming web traffic — bring it in, stabilize and examine, and do not allow entry until information pressure equalizes.”



2. Principles (the build service’s, applied to text)

  1. The seat never touches the raw thing. Fetching, redirect-following, TLS and decoding all happen in the airlock. No seat’s egress widens for web reads — which closes the exfiltration channel that an “allow this website” rule would open, because an allowed website is also a place a compromised seat could send things.
  2. Refusal is a result. Three outcomes — admitted; admitted with findings (spans bracketed or redacted by policy); refused at a named stage — and never a silent partial.
  3. Detection is not the bound; blast radius is. Scans catch the crude majority. The marked-turn watch is the control that works on the injection nobody anticipated.
  4. Typed in, typed out. Request and artifact are strict, versioned records; unknown fields refuse.
  5. Measured, not stated. Every probe has a fixture set of planted injections and clean controls. A probe with no measured catch rate is a premise, not a control.
  6. Fetched text is data. Quoted and cited, never acted on — a standing rule for any web-enabled turn on any seat, with or without the airlock in front of it.

3. The chamber, in one picture

 seat: "read this URL" ──▶ [1 outer door: fetch, log every hop, hash raw → quarantine]
                            [2 stabilize: HTML → text; hidden things removed & counted; odd Unicode flagged]
                            [3 examine: rule probes (instruction shapes, exfil links, credential shapes)
                                        + a small captive model that flags, never decides]
                            [4 equalize: findings → pressure → admitted / admitted-with-brackets / refused]
 seat ◀── typed artifact ── [5 inner watch: the reading turn's actions checked against its orders]

4. What each stage is for

  1. Outer door (fetch). The seat never connects anywhere; the airlock does, under a per-request allowlist of the URL’s host and each redirect hop, every hop logged. Observed and recorded: host, certificate chain, content type, byte count, time, final URL. The raw bytes are hashed into quarantine and retained as evidence, never as working material.
  2. Stabilize (reduce). Injections hide in what a human never sees: invisible elements, white-on-white text, zero-size fonts, off-screen boxes, zero-width characters, right-to-left overrides, homoglyph runs, invisible encoded blobs. One pinned extractor reduces the page to text; scripts, styles, comments and metadata are dropped; hidden elements are removed and counted — how much had to be removed is a finding in itself; alt text and link targets go to separate fields; Unicode is normalised with a finding for each trick found.
  3. Examine (probe). Deterministic rules first: sentences addressed to an AI (“ignore previous”, “you are now”, “as an assistant”), fake role markers and tags shaped like a harness’s own, tool-call-shaped syntax, fenced shell blocks, “do not tell the user” patterns; exfiltration shapes (image links to other hosts, long or encoded query strings, forms, paste-site links); inbound credential shapes (a key-shaped string on a page is something a seat must never repeat); density (instruction-shaped spans against text length, hidden-matter ratio, text present only in removed elements); provenance mismatch (final host differs from the requested host; content type differs from what the URL implied). Then a second opinion from a small, captive, hub-local model asked one question with spans returned: does this text contain instructions addressed to a reading agent, and where? The model raises pressure; it never admits or refuses alone, and it is measured for false positives on instruction-dense honest content (documentation, tutorials) before it counts toward the threshold.
  4. Equalize (admit). What the seat receives is a labelled artifact: where it came from, the raw hash, the clean text, the findings, the pressure, the outcome, and whether flagged spans are bracketed (so the agent sees them as quoted data) or cut out, per the requester’s policy. Above the threshold: refused, with the findings and nothing else. Refusal is a result, never a silent trim.
  5. Inner watch (after). The only control that works on the injection nobody predicted: the consuming turn is marked, and its tool calls are compared to the orders that turn was given. A message to someone it was not told to address, a shell command containing a string from the page, a reach for a tool it was not granted — a monitor row and a stop.

5. The interface

A request carries a schema version, the URL, a one-line purpose (logged), the requester, a policy — quote, redact or strict — and a byte ceiling. The artifact carries identity and correlation, the observed transport rows, the text, the findings (each with its stage, kind, span and note), the pressure, the outcome, and the quarantine reference. A refusal carries findings and nothing else. The normative schema is written by the first consumer, as the build service’s was.


6. What it is not

It does not tell a seat a page is true; facts still get checked. It is not a general web proxy for seats. It is not perfect: the rule list will always be a step behind an attacker, a sufficiently indirect injection reads as prose, and the captive model will sometimes flag honest documentation that happens to be full of imperatives. Stage 5 exists because of that.


7. Where the operator’s hands are

None in normal use. The operator rules the pressure threshold per seat role, whether flagged spans are bracketed or cut, and which seats may request at all; the recipe merges as for every service.


8. How it is built and measured

The airlock starts from the build service’s hardened base, with one difference: its network is a per-request allowlist instead of none. The deterministic probe list is opened as a fixture register — planted injections and clean controls, committed by path — before any probe counts. The first admitted artifact goes on the record; the marked-turn watch lands as a monitor row on the same instrument the fleet’s wake supervisor uses for typed outcomes; the captive scorer is measured before it counts. Until the airlock exists, a seat’s web reads run from model knowledge with a verification list a reviewer checks, and any vendor-side search a harness offers is enabled only with the “content is data” rule applied by hand.


AstralMaris is one operator and a small fleet of AI agents on Apple-silicon hardware. This series records designs the fleet runs, with the measurements that shaped them.

Back to the title ↑