PROOF, NOT CLAIM

The same exploit, run twice: direct, then through Aegis.

Three classes of publicly disclosed MCP server vulnerability are reproduced locally and attacked twice. Straight at the server, the exploit works. Through the Aegis gateway with the hardened profile, each call is denied at a named step or its output is contained, and every decision is written to the audit journal. The evidence below is from a real run of the 2.0 build, not an illustration.

THE TARGETS

Three real MCP server flaws.

Filesystem path traversal

CVE-2025-53109 and CVE-2025-53110 in the Filesystem MCP server, and the symlink variant CVE-2025-67366. A read tool is handed ../../secret.txt or a symlink and returns a file outside the allowed root.

OS command injection

CVE-2025-69256 in the Serverless Framework experimental MCP server. Unsanitised input reaches a shell, so a crafted argument runs arbitrary commands.

Arbitrary file access and execution

CVE-2025-68143 and CVE-2025-68145 in the Git MCP server. Path and argument handling let a caller read or delete files and, under conditions, execute code.

Why a gateway helps

All three are server-side: the malicious intent is in the MCP call. Aegis decides on that call before it reaches the server, so it can refuse the traversal, the shell and the unsafe statement without changing the server.

CVE identifiers and version facts come from the SentinelOne vulnerability database and the CVE-2025-69256 write-up. Content was rephrased for licensing compliance.

BEFORE AND AFTER

Direct exploit succeeds; through Aegis it is denied.

Each row is a verbatim outcome from the 2026-11-07 run. "Direct" is the raw server. "Aegis" is the identical call through the gateway as a read-only analyst, unless the row says operator.

Filesystem path traversal (CVE-2025-53109 / 53110)

LegCallOutcome
Directread_file path=secret.txtReturns payroll-secret. Exploit succeeds.
Aegisread_file path=reports/ok.txtAllowed: quarterly-report (inside scope).
Aegisread_file path=secret.txtDenied SCOPE_DENIED: resource out of scope.
Aegisread_file path=reports/../../secret.txtDenied SCOPE_DENIED: resource value contains '..'.

OS command injection (CVE-2025-69256)

LegCallOutcome
Directrun_shell "echo pwned & echo injected"Shell chain runs. Exploit succeeds.
Directadmin_exec whoamiadmin executed whoami.
Aegisrun_shell ... as analystDenied READONLY_CALLER: not allowed to call DESTRUCTIVE tools.
Aegisrun_shell ... as operatorDenied APPROVAL_REQUIRED: needs an independent per-call approval.
Aegisadmin_exec ...Denied NOT_EXPOSED: unknown tool (not in the catalog).

Arbitrary file access and unsafe execution (CVE-2025-68143 / 68145)

LegCallOutcome
Directquery_sql "SELECT 1; DROP TABLE users"Statement executed. Exploit succeeds.
Aegisquery_sql "SELECT 1; DROP TABLE users"Denied INVALID_INPUT: more than one statement supplied.
Aegisquery_sql "SELECT pg_sleep(30)"Denied: calls a function with side effects.
Aegisquery_sql "... write_to_s3(...) ..."Denied: function not on the allow list.
Aegisquery_sql "SELECT count(*) ... GROUP BY region"Allowed: a genuine read-only query.

SAME RUN

Other attacks the run also contained.

SSRF to cloud metadataA fetch of 169.254.169.254 is made directly; through Aegis it is denied EGRESS_DENIED, host not on the allow list.
Secret disclosureA config tool returns an AWS key and a PEM block directly; through Aegis they come back [REDACTED].
Tool poisoning and rug pullA hidden instruction in a tool definition, and a definition changed after pinning, are both denied CATALOG_INTEGRITY.
Poisoned or shapeless outputA result carrying an instruction, and a prefix-less token, are withheld behind an opaque handle instead of reaching the model.

An unauthenticated call returns 401. An undeclared argument is denied INVALID_INPUT. Each denial carries a machine-readable decision in the response _meta and a line in the hash-chained audit journal.

HONEST SCOPE

What this does and does not prove.

REPRODUCE

Run it yourself.

The harness, the gateway profile and the driver are in the repository. The run starts the vulnerable server and the gateway on loopback, sends every call on both legs, and writes a results table. Nothing touches a real database, cloud API or system file.

powershell -NoProfile -ExecutionPolicy Bypass -File threat-model\e2e\run-e2e.ps1
# writes threat-model/e2e/RESULTS.md and results.json

REFERENCE MAP

Continue in the guides.