orca/docs/reference/windows-edr-posture.md

Windows EDR signal surface

Orca's Windows process tree is shaped like the thing behavioural EDR is built to find. An enterprise Windows 11 / Intune tenant opened six Microsoft Defender for Endpoint incidents against Orca 1.4.192 in eight days. All six fired as active incidents and stayed open; three closed only because a human classified them by hand in the portal. Defender never downgraded or closed one on its own.

None were signature hits. Every one was behavioural process-tree scoring, and two escalated to multi-stage incidents carrying ATT&CK tactic mappings (Execution, Collection).

The framing this document keeps throughout, because both halves matter:

Defender is not malfunctioning. It is describing the code accurately. Orca really does copy its own signed image under a different name, really does read every process's memory on a timer, really does run base64-encoded PowerShell with the execution policy bypassed, and really does take screenshots and synthesise input from a runtime-compiled assembly. Each of those is a deliberate engineering choice with issue history behind it. The problem is not that the capabilities are illegitimate — it is that their behavioural signature overlaps with attack techniques, and an EDR scoring behaviour cannot see the difference.

Do not read this as a bug report against Defender, and do not read it as a claim that Orca is malware. It is a map of which of our behaviours are legible to an EDR as attack-technique-shaped, why each one exists, and what engineers and administrators can do about it.

What the tenant actually saw

Four independent evidence clusters, from six incidents:

Cluster	Incidents	Evidence
Update	A, B, C	orca-windows-setup.exe → old-uninstaller.exe, Uninstall Orca.exe (electron-builder generates these; they are in no repo file)
Spawn	all six	Orca.exe → orca-terminal-daemon.exe → powershell.exe / pwsh.exe / cmd.exe / reg.exe → claude.exe, gh.exe, codex.cmd
Process table	D	"suspicious memory activity" — OpenProcess plus a PEB read against every process on a repeating cadence
Computer use	E, F	runtime.ps1, computer-sidecar.js, many operation.json, a burst of ~10 short-lived powershell.exe

Incident E is the one to look at hardest: 5 alerts, 37 evidence items, ATT&CK Execution + Collection, and a description reading "Screenshots were taken unexpectedly on this device… Screen capture code was found in a script launched by powershell.exe." Incident F added "suspicious MSIL code", from the Add-Type -TypeDefinition that recompiles inline C# P/Invoke on every operation.

In the update cluster the uninstaller is genuinely NotSigned, while Orca.exe and orca-terminal-daemon.exe report Valid CN=SignPath Foundation.

The behaviours, and why each one exists

The daemon runs from a copy of our own image

src/main/daemon/daemon-host-relocation.ts copies the Electron runtime into %LOCALAPPDATA%\Orca\daemon-host\<version>\ and forks the terminal daemon from there.

It exists because the NSIS installer deletes the old install directory and force- kills every process imaged under it. Without relocation, an auto-update kills the terminal daemon and every live terminal with it. The copy is a run-as-node Orca.exe rather than node.exe so there is no console flash and asar still resolves; config/nsis/orca-installer-hooks.nsh reaps it on a real uninstall (guarded by ${isUpdated} so an update's uninstallOldVersion never fires it).

At the time of these incidents the copy was also renamed to orca-terminal-daemon.exe, the image name every incident here reports, and DAEMON_HOST_EXE_NAME's comment stated the reason without varnish: "so the NSIS updater's taskkill /IM Orca.exe can't match it." The rename has since been removed; the copy now keeps the app exe's own file name, because the updater's kill sweep is path-scoped on every host that has PowerShell and the rename only ever bought the no-PowerShell fallback. See windows-daemon-host-relocation.md.

How an EDR reads it: MITRE T1036, masquerading — and, for what remains, T1036.005. A signed executable copied out of the install directory into %LOCALAPPDATA% under a different name, which then spawns shells, matches the textbook description closely enough that no behavioural engine can be expected to score it low. Dropping the rename removes that literal indicator but not the underlying shape: execution from a non-standard user-writable location is scored on its own. Note also that the strongest form of the T1036 signal was never present here — the shipped binary's OriginalFilename is empty, so there was no embedded name for the old disk name to contradict.

Every process gets a handle, on a timer

src/main/windows/windows-process-table.ts takes a Toolhelp32 snapshot under one flag set, CommandLine | CreationTime, shared by every caller. pid, ppid and name come out of the snapshot itself and open nothing. CommandLine is what opens a handle: the addon calls GetProcessCommandLine per process, which opens PROCESS_QUERY_INFORMATION | PROCESS_VM_READ and walks the PEB with three ReadProcessMemory calls (src/process_commandline.cc:32,41-47 in the vendored @vscode/windows-process-tree 0.8.0 source that config/patches/ patches).

Memory is retired as of this change, and that is a real reduction: it made GetProcessMemoryUsage open a second PROCESS_QUERY_INFORMATION | PROCESS_VM_READ handle per process for a GetProcessMemoryInfo call whose result no caller read (src/process.cc:47-63). Dropping it halves the handles opened per snapshot. It does not remove the remote memory read, because the command line still performs one.

It exists because seven independent readers used to fork powershell.exe for a Get-CimInstance Win32_Process scan. That cost, measured: a PowerShell Transcription policy recorded ~289 GB across 1.4 million files because a scan ran every ~2 seconds (#15209); a Group Policy or AV block turned a query into "unavailable", which callers read as "no evidence", which is how a PTY tree survived its own teardown (#9045, #10475); and the scan cost ~700 ms per pane, so panes multiplied it (#15036). The native snapshot answers the same question in 15.9 ms against 706 ms for CIM — p50, measured on Windows 11 at 1050 processes. See windows-process-enumeration.md.

Asking for fewer fields is cheaper, and the module now asks for the smallest set that still answers every caller. There is no per-flag-set cache split: one TTL-cached snapshot serves everyone, deliberately, because a split would restore the per-pane fan-out the cache exists to remove — a 32-wide teardown has to collapse into one scan. So the cheap identity-only read is not something any caller can select; every read pays for CommandLine. An earlier revision of this file described a two-cache design with 6.3 ms / 12.3 ms p50 figures at 492 processes. That design is not in the tree and those numbers describe no code path here; the figures that do apply are the module's own, in windows-process-enumeration.md.

How an EDR reads it: a cross-process handle plus a remote memory read against every process on the box, repeating on a cadence, is the read half of the telemetry that credential dumping and process injection produce. MDE surfaced it as "suspicious memory activity".

That signal is still present. An earlier revision of this file claimed the command line "now comes from the kernel" through NtQueryInformationProcess's ProcessCommandLineInformation class, needing only PROCESS_QUERY_LIMITED_INFORMATION, and that ReadProcessMemory was absent from the compiled addon. None of that is true of the code we ship. process_commandline.cc calls NtQueryInformationProcess with ProcessBasicInformation only — to locate the PEB — and then issues three ReadProcessMemory calls against a PROCESS_VM_READ handle to read the PEB, the RTL_USER_PROCESS_PARAMETERS, and the command-line buffer. Nothing asserts an import table, and no such assertion would pass.

What this change did remove is the Memory flag's second handle and its GetProcessMemoryInfo call, so the per-process handle count per snapshot halves. What remains to declare to administrators is unchanged in kind: one PROCESS_QUERY_INFORMATION | PROCESS_VM_READ handle and a PEB read against every process on the box, at the shared snapshot's cadence. Moving to ProcessCommandLineInformation (Windows 8.1+, PROCESS_QUERY_LIMITED_INFORMATION only) would genuinely retire the remote read, but it is an addon patch nobody has written; treat it as unclaimed work, not as shipped.

Encoded, policy-bypassing PowerShell

Three sites are named in the incident analysis:

• src/relay/windows-port-scan.ts ran -NoProfile -NonInteractive -ExecutionPolicy Bypass -EncodedCommand over a Get-NetTCPConnection -State Listen script to find dev-server ports. Enumerating listening ports is MITRE T1049, network service discovery, and doing it through an encoded policy-bypassed shell is the aggravating factor rather than the finding itself. The ordinary scan now starts no PowerShell at all — netstat.exe -ano, with the owning process name projected off the shared native table — and that payload survives only as the last-resort fallback, as -Command with no policy override.
• src/main/daemon/shell-ready.ts uses -EncodedCommand for the OSC 133 bootstrap.
• src/main/agent-hooks/windows-powershell-hook-launcher.ts wraps managed hooks.

No site spells the pair any more. src/main/ssh/ssh-remote-powershell.ts, src/shared/setup-agent-sequencing.ts, src/shared/windows-cmd-runner-delayed-launch.ts and src/shared/windows-interactive-login-spawn.ts each dropped -ExecutionPolicy Bypass as a measured no-op: the policy gates script files, never -EncodedCommand. Where the bypass was load-bearing it moved in-payload as a process-scope Set-ExecutionPolicy (setup-agent-sequencing.ts), which is the pattern to copy rather than restoring the switch — the switch loses to a GPO scope anyway, so it never covered the locked-down case.

What remains is -EncodedCommand without the bypass: the PTY bootstraps (src/main/daemon/shell-ready.ts, src/main/providers/local-pty-shell-ready.ts, src/main/providers/windows-shell-args.ts), the hook wrappers (src/main/agent-hooks/windows-powershell-hook-launcher.ts and its callers src/main/agent-hooks/runtime-home-hook-command.ts, src/main/agent-hooks/installer-utils.ts, and src/main/claude/hook-settings.ts — that last one only as a fallback since #18875, see below), src/main/runtime/windows-default-route-interfaces.ts, src/main/runtime/orchestration/setup-completion-signal.ts, src/shared/hermes-startup-query.ts, and the four ex-bypass sites above. src/main/runtime/windows-mobile-firewall.ts encodes a script and launches it elevated through Start-Process -Verb RunAs, which is a stronger shape than any of those; only that hop is encoded, because -ArgumentList re-splits an unquoted parameter string on whitespace.

One site still spells -ExecutionPolicy Bypass with no encoding, the weaker signal: src/main/cli/wsl-cli-scripts.ts (-File, and it is a real script file, so the switch is not a no-op there). src/main/system-fonts.ts dropped it for plain -Command; src/shared/secure-path-windows-acl.ts no longer runs PowerShell at all, having moved to icacls.exe; and computer use now asks for -ExecutionPolicy RemoteSigned in src/main/computer/windows-powershell-execution-policy.ts, falling back to Bypass only after a policy-blocked start.

Regenerate with rg -- '-EncodedCommand|-ExecutionPolicy' src/ rather than trusting the lists above, and note that a raw grep under-reports: the hook sites reach -EncodedCommand through wrapWindowsPowerShellEncodedCommand and never spell the flag themselves.

Encoding is not gratuitous: it shields paths and switches from cmd.exe and MSYS rewriting (#6078, #14815), which is a real class of corruption. But -EncodedCommand is a first-class Defender alert title ("Suspicious PowerShell command line"), and base64 raises the score rather than lowering it, because it denies the analyser the payload it would otherwise clear.

The hook launcher is prior art worth knowing about. #16003 measured, on a reporting Kaspersky host, that -WindowStyle Hidden paired with -EncodedCommand was denied at CreateProcess with exit 126 regardless of payload — exit 0 was denied too. The fix was to stop spelling the flags: WINDOWS_POWERSHELL_HOOK_SWITCHES is now just -NoProfile, and separately, in #16576, the execution policy bypass moved in-payload as a process-scope Set-ExecutionPolicy — a real command-line signal reduction, though #16003's measured denial keyed on -WindowStyle Hidden + -EncodedCommand, not on the bypass. It is also honest that the underlying behaviour did not change.

Copy the pattern, but copy its caveat too. windows-powershell-hook-launcher.ts records that dropping -WindowStyle Hidden was a real tradeoff whose suppression "was never measured" and "remains unverified on a real box". Reducing spelled flags is the right instinct; treat any specific claim about what a removed flag was doing as unproven until someone measures it.

cmd.exe /c carrying caret-escaped free text

buildWindowsCmdShimCommandLine in src/shared/child-process/windows-command-line.ts builds /d /v:off /s /c "…" for the .cmd and .bat targets Windows can only start through cmd.exe (codex.cmd being the one that matters). Because cmd expands %VAR% even inside a quoted token, each % is broken with "^%".

The escaping is not decorative. Measured on Windows 11 against a real .cmd shim, ["a b", 'c"d', "e%F%g", "h&i", "j^k"] came back as ["a b", 'c"d', "e^%F^%g", "h"] — the & truncated the argument and ran the remainder as a command.

How an EDR reads it: caret escaping is the canonical obfuscation marker in cmd.exe command lines, and the free text being escaped here is an agent prompt, so the line is long, high-entropy, and attacker-shaped. It is the exact input an obfuscated-command-line detector is tuned on.

The spawn tree itself

Orca.exe → the relocated daemon host (orca-terminal-daemon.exe in the builds these incidents cover, Orca.exe since) → a shell → an agent CLI is what a terminal multiplexer for coding agents is. reg.exe appears from src/main/win32-utils.ts, src/main/agent-hooks/managed-hook-owner-identity.ts and src/relay/pty-shell-utils.ts (reading the OpenSSH DefaultShell).

Nothing here is avoidable in principle. What is controllable is depth and breadth: every interpreter hop between Orca and the thing the user asked for adds a scored edge, which is why the shipped doctrine of #15520 and #15595 is to shorten the interpreter chain rather than to hide a window.

#18875 is a worked example of that doctrine. The Claude Code lifecycle hook was registered as powershell.exe -NoProfile -EncodedCommand <...> whose entire decoded payload was a Test-Path and a call to ~/.orca/agent-hooks/claude-hook.cmd. It now registers the script path itself (<path> || echo {}), so bash -> powershell -> cmd -> curl became bash -> cmd -> curl and one powershell.exe -EncodedCommand per hook event — a first-class Defender alert title — leaves the tree. The reporting box fired ~6 900 of them in five days, 70% from Claude sessions that were not running under Orca at all and whose hook exits at its first ORCA_PANE_KEY guard.

What is measured is latency and the hop count, nothing else: median 471 ms -> 213 ms per event idle, and 656 ms -> 296 ms (p95 696 ms -> 337 ms) under 10-way concurrency, invoked as Claude Code invokes it. No EDR verdict on either tree was measured, so claim the removed -EncodedCommand spelling and the shorter chain, not a score. cmd.exe remains in the tree, spelled by MSYS's own .cmd spawn rather than by us — the doc's one "unavoidable for .cmd/.bat" case, carrying an absolute path and two literal tokens, with no caret escaping, no encoding and no free text. The encoded launcher is still the shape for profile paths the shells cannot carry bare (a space, %, ^, &, non-ASCII, a UNC profile) and for hosts where Git Bash is not resolvable, because PowerShell 5.1 rejects || (measured: parse error, exit 1).

That last clause is the standing assumption of this change, and it is worth stating plainly because it is not measured. || parses in Git Bash, cmd.exe and pwsh, but not in Windows PowerShell 5.1, so the direct shape is correct for any host that is one of the first three. Claude Code itself is a Git Bash host on native Windows. What no one here has verified is which host a compat consumer uses: cursor-agent and Devin import ~/.claude/settings.json and run command through their own launcher (the managed .cmd carries a DEVIN_PROJECT_DIR skip for exactly that). If one of them spawns hook strings through Windows PowerShell 5.1, its imported Claude events become a parse error with empty stdout, which is the fail-closed case #14818 exists to prevent. The encoded launcher had no such assumption — it was a powershell.exe invocation and therefore parsed anywhere. Before widening the direct shape to another agent, measure that consumer's host.

Computer use: screen capture, synthetic input, runtime-compiled MSIL

native/computer-use-windows/runtime.ps1 is a large PowerShell script. src/main/computer/desktop-script-provider-bridge.ts launches it as powershell.exe -NoLogo -NoProfile -NonInteractive -ExecutionPolicy RemoteSigned -File runtime.ps1 <operation.json>, retrying once at Bypass only if the start comes back policy-blocked — once per operation, with desktop-script-provider-client.ts writing a fresh operation.json into a new temp directory each time. On every launch the script runs Add-Type -TypeDefinition over inline C# that P/Invokes SendInput and the window APIs, then captures the screen through Graphics.CopyFromScreen.

That is four separate high-signal behaviours stacked in one process:

Behaviour	How it is scored
Graphics.CopyFromScreen	MITRE T1113, screen capture — Collection tactic
SendInput synthetic keyboard/mouse	input synthesis against other applications
Add-Type -TypeDefinition on every operation	MSIL compiled at runtime; incident F's "suspicious MSIL code"
One powershell.exe per operation	a burst of short-lived interpreters under one parent

The bottom two rows are the two the incident text named directly, and they are also the two a persistent runtime host would remove: a long-lived helper compiles its P/Invoke stubs once and answers operations over a channel, so neither the MSIL recompilation nor the interpreter burst repeats. A change doing that is in flight and unmerged at the time of writing; check the code rather than this paragraph for what the shipped build does. Screen capture and SendInput are inherent to the feature and no refactor removes them.

Signing is not the gate

The most useful calibration in the whole incident set came from the reporter's own machine: Antigravity IDE's main executable is NotSigned and was not flagged, while Orca's is signed and was flagged six times. Their conclusion: "signing is not the gate here — behaviour is."

The mechanism is that Defender reputation is signer plus prevalence, and prevalence is keyed on file hash. A widely installed unsigned binary clears on install count alone. Orca's signature is a free OV certificate from SignPath Foundation (config/electron-builder.config.cjs sets win.signtoolOptions.publisherName; config/scripts/verify-windows-inner-signature.mjs pins CN=SignPath Foundation, O=SignPath Foundation, L=Lewes, S=Delaware, C=US), shared across many OSS projects, with no independent SmartScreen or MAPS reputation of its own. Every release ships new hashes, so whatever prevalence a build accumulates resets on the next update. Dev channels ship unsigned by design, because SignPath's approval waits cannot fit a dev cadence (config/scripts/verify-dev-channel-packaging.mjs).

Signing the uninstaller is worth doing — an unsigned old-uninstaller.exe running under a signed installer is a gratuitous contribution to the update cluster — but do not expect it to change the behavioural verdict. The three non-update clusters contain no unsigned binary at all.

What we do not know

Two limits the incident analysis recorded, kept here rather than smoothed over:

• No data on Hermes. Nothing in this document describes how Hermes behaves under the same tenant policy — though src/shared/hermes-startup-query.ts does spell -EncodedCommand, so the gap is telemetry, not surface.
• Antigravity not being flagged is absence of evidence, not proof. It is one reporter's recollection from one machine, not a measurement. It is strong enough to falsify "the problem is that we are not signed well enough"; it is not strong enough to support a positive claim about how Defender scores that product.

Add to those: this is one tenant with one policy configuration. Whether the same build scores the same way elsewhere is unmeasured.

Guidance for engineers

Fixes for several of the shapes above are in flight in separate changes; nothing in this section should be read as a statement that a given site has already changed. Check the code before relying on it.

The checklist. On Windows, do not reach for:

Don't	Instead
-ExecutionPolicy Bypass on the command line	Set the policy in-payload at process scope, as windows-powershell-hook-launcher.ts does, or do not run a .ps1 at all
-EncodedCommand	A temp .ps1 with an argument, or no PowerShell hop: prefer a native API or an existing Node path
cmd.exe /c carrying escaped free text	Spawn the real target directly. cmd.exe is only unavoidable for .cmd/.bat; keep free text out of the line where you can
Forking powershell.exe to read system state	The native reader — windows-process-enumeration.md is the standing rule for the process table
A process per operation in a loop	One long-lived helper with a request channel. A burst of short-lived interpreters under one parent is itself the signal
Add-Type -TypeDefinition at runtime	A precompiled, signed assembly, or a native helper
Copying our own image under a different name	Copy it verbatim — windows-daemon-host-relocation.md (done for the daemon host)
Deriving a script runner from a UI preference	windows-setup-shell.md — the script declares its own interpreter

Two framing rules that outlast the table:

• Shorten the interpreter chain. Each hop between Orca and the user's actual target is a scored edge and a place for AV to deny a CreateProcess. This is the shipped doctrine of #15520 and #15595.
• Do not spell a flag you can avoid spelling. #16003 measured a denial that was independent of the payload and keyed purely on the switch combination on the command line. What is on the line is itself the detection surface.

Guidance for administrators deploying Orca

Path exclusions alone will not silence these

This is the single most important operational point, and it is the one most commonly got wrong. The six incidents are MDE EDR behavioural alerts. Defender Antivirus path exclusions suppress scan detections; they do not suppress EDR behavioural alerts the same way. Adding %LOCALAPPDATA%\Programs\orca\ to the AV exclusion list and expecting the incidents to stop will not work.

What actually stops incidents being created

An MDE alert suppression rule scoped to the process tree. Build it in Microsoft 365 Defender (Settings → Endpoints → Alert suppression), conditioned on:

• Alert titles — A suspicious file was observed and Suspicious PowerShell command line, plus any further titles your tenant actually produced. Take the titles from your own incidents rather than from this list.
• File paths — Orca.exe and orca-terminal-daemon.exe under %LOCALAPPDATA%\Programs\orca\ and %LOCALAPPDATA%\Orca\daemon-host\.

Scope it as narrowly as your tenant will tolerate, and review it when Orca updates: the daemon-host path carries a <version> segment, so a rule pinned to one version will silently stop matching. Two traps in that path in particular. Materialization stages into a <version>.staging-<hex> sibling before renaming it into place, so an exact-version rule misses the tree mid-update — which is precisely when the update-cluster incidents fire. And the root falls back to the Electron userData path when LOCALAPPDATA is unset, so %LOCALAPPDATA%\Orca\daemon-host\ is the normal location rather than a guaranteed one. Prefer a prefix match on …\Orca\daemon-host\ over a rule pinned to one full path.

Add AV path exclusions for those two directories as well — they cut scan cost on a tree that is rewritten on every update — but understand the division of labour. The exclusions reduce scanning; the suppression rule is what stops incidents being created.

Check your ASR rules

Check whether the tenant has the Attack Surface Reduction rule "Block executable files from running unless they meet a prevalence, age, or trusted list criterion" enabled. If it is, that alone explains a freshly signed Orca build being hit immediately after every update: each release ships new hashes, so every build starts at zero prevalence and zero age no matter how it is signed. Either allowlist the Orca install paths for that rule or expect a hit on each update.

Expect the alerts to recur after each update

Prevalence is keyed on file hash. An update replaces the hashes, the reputation starts over, and a suppression rule is the only thing carrying across.

Computer use: decide before you deploy

Read this section before enabling computer use on a monitored endpoint, not after.

On a monitored endpoint, an alert reading "Screenshots were taken unexpectedly on this device" is not the kind of finding a SOC dismisses on sight.

Incident E is the shape to expect: 5 alerts, 37 evidence items, a multi-stage incident mapped to ATT&CK Execution + Collection, and a description naming screen capture found in a script launched by powershell.exe. Incident F adds runtime-compiled MSIL to the same tree.

Every part of that is an accurate description of what the feature does. Orca's computer use takes screenshots, synthesises keyboard and mouse input into other applications, and compiles the P/Invoke stubs it needs at runtime. An organisation that monitors for Collection-tactic activity — and any organisation running MDE with default incident creation does — will see it, and will see it as Collection.

So decide deliberately, in advance:

• Allowlist it, with a suppression rule covering the computer-use tree (powershell.exe with -File …\runtime.ps1) as well as the base Orca paths, and tell your SOC what it is before the first incident rather than during it.
• Or leave it disabled on monitored endpoints.

What does not work is deploying it un-triaged and handling the incidents reactively. By the time a Collection-tactic incident is open, an analyst is already reading a description of screenshots being taken without the user's knowledge, and the burden of proof has moved to you.