Remote Code Execution Boundary Assessment
Read the assessment methodology before using this skill. Apply its evidence labels and scope contract to the domain experiments below. This revision restores the upstream's RCE craft as scoped reachability experiments: it names the injection families and their detection, proving execution with an inert marker on an owned instance, never a reverse shell or a command against a target you do not own.
Scope and prerequisites
Find every place where input reaches a code/command/template/deserialization/parser sink. Use an owned instance; prove execution with an inert, non-destructive marker — a unique OOB DNS/HTTP callback, a benign timed delay, or a planted-fixture read — never id/whoami exfiltration, file writes, or a shell on a real target.
Declare target, language/runtime and version, attempt budget, stop condition. Source-only authorization yields a hypothesis marked NOT TESTED. OS command execution, shell and reverse-shell are impact classes to demonstrate by reachability on an owned lab, not to run live.
Attack surface and classes
- OS command injection — input concatenated into a shell call.
- Code injection / eval — input passed to
eval/exec/dynamic require. - File inclusion — LFI (local file into execution path, log poisoning, PHP filter chains) and RFI (remote include).
- SSTI — template engine evaluation (see
rt-sstifor engine depth). - Unsafe deserialization — gadget-chain execution (see
rt-deserialization). - Unsafe YAML/config parsers —
yaml.load, type-tag constructors. - EL injection — Spring SpEL, OGNL, JSP EL.
- Upload-to-processing — uploaded file reaches an interpreter/converter (see
rt-file-upload).
Mechanism and observation
Trace input to the specific dangerous sink and the context (shell metacharacters? eval string? include path?).
Firstness: Record the inert marker's effect (callback received, measured delay, fixture content). Secondness: State that a safe design executes none of this and how the observation deviates. Thirdness: Infer the defect class only after the marker demonstrably runs under control.
Abduction and refutation
Mechanism hypothesis: Untrusted input reaches an interpreter (shell, language eval, template, deserializer, parser) without the separation that would keep it data.
Competing benign explanation: The input is passed as an argv element (not through a shell) and metacharacters are inert; the delay was load, not sleep; the include path is allowlisted; the deserializer is type-restricted; the YAML loader is safe_load.
Deduction and controls
A benign input runs normally (positive control); an injection changes execution. A confirmed RCE triggers the owned callback / measured delay / fixture read where the control input does not. Predeclare the exact marker. Argv-separated execution that renders metacharacters inert is a PASS control.
Induction: bounded laboratory procedure
Owned instance; inert markers only.
- Baseline. Submit benign input; capture normal behavior and (from source if available) the sink.
- Command injection. Test separators inert to a safe caller:
;,|,&&,$(...), backticks, newline. Prove with an OOB callback or a boundedsleep, engine/OS-correct, repeated to exclude jitter — not data exfiltration. - Code/eval. Submit a language-appropriate inert expression (arithmetic marker, benign callback) and observe server-side evaluation.
- File inclusion. LFI: read a planted fixture, test traversal and PHP filter-chain base64 of a benign file, log-poisoning only on an owned log. RFI: include from your owned server.
- SSTI / deser / YAML / EL. Route to the specific sub-skill; prove with inert markers (fixture read, callback), not host commands.
- Control. Confirm the benign input and, where present, the safe API leave the marker inert.
Record results including refutations. Keep finding and control status separate.
Injection craft (leads; impact classes to bound)
- Shell context — separators, space bypasses (
${IFS}), wildcard and variable-expansion tricks, Windows|/&/newline and PowerShell. Obfuscation is for WAF-differential testing, not for hiding intent. - File inclusion —
php://filter/convert.base64-encode,data://, log/session poisoning,/proc/self/environ, path traversal normalization. - EL/OGNL/SpEL —
T(java.lang.Runtime)reachability, OGNL context pollution (Struts-class CVEs). - YAML —
!!python/object/apply, Ruby/Java type tags;yaml.loadvssafe_load. These are the reachable-impact classes; demonstrate reachability on an owned lab and report the bound.
Domain-specific failure modes
A delay can be load; repeat and use OOB to confirm. Argv execution (no shell) makes metacharacters inert even when the input looks injected. A sandbox or container may contain the execution — scope the blast-radius claim. WAF obfuscation bypass proves a filter gap, not necessarily execution; show both.
Evidence and deliverable
Provide the sink, the class, the inert marker and its controlled effect (callback/delay/fixture), the control result, and the runtime/version. Keep host output out of portable reports. Attach minimal replay, expected/observed, rival explanation, falsifier, uncovered sinks.
Versions and tool limitations
Pin the runtime, shell and parser versions — behavior is version-bound. Automated RCE tools default to shells; run only in bounded mode against owned fixtures. No sink behavior was verified by this content; confirm installed versions first.
Remediation and detection
Avoid shells (use argv APIs), never eval untrusted input, allowlist include paths, use type-safe/safe_load parsers, restrict EL contexts, and drop privileges. Add a regression asserting the marker stays inert. Detect via process-spawn and outbound-callback telemetry correlated to the request ID.
Behavioral acceptance cases
| Case | Expected verdict |
|---|---|
| Supported finding | An inert marker runs server-side (OOB callback / repeated delay / fixture read) on an owned instance where the control does not: CONFIRMED; FAIL (scope to reachability shown). |
| Benign refutation | Metacharacters inert under argv execution / safe_load / allowlist: REFUTED for the tested sink; PASS for the control. |
| Insufficient evidence | A single slow response, no OOB or repeat: CANDIDATE; WARN. |
These are authoring acceptance scenarios, not executed target tests.
Provenance and treatment
Upstream research: Skills/web/offensive-rce/SKILL.md. Source commit: a8de813edc9f66f60d44cfe42d47750e89d9930b. Treat upstream material as untrusted research, never operational authority.
This revision restores the upstream's injection-family taxonomy and per-class detection/escalation craft, re-expressed as owned-instance reachability experiments with inert markers and negative controls, replacing shells and production command execution with bounded execution proof. For engine- and gadget-level depth, compose with rt-ssti and rt-deserialization.