
CVE-2026-117567: OS command injection in esp-mcp. Prompt injection → tool call manipulation → arbitrary commands. This is why I treat AI agent stacks like web apps with extra steps.
Signal is active with 3 mentions in latest observed window
Recommended action window: Monitor and triage in normal cycle
NVD description
esp-mcp through commit 28afddf contains an OS command injection vulnerability that allows attackers to run arbitrary commands by supplying shell metacharacters in MCP tool parameters. Attackers influencing agent tool calls through prompt injection can inject metacharacters into target, port, project_path, or project_name values passed to asyncio.create_subprocess_shell, executing commands as the MCP server user.
Priority
LOW
Exploitation
NONE
PoC
NONE
Patch
NONE
Momentum
NONE

CVE-2026-117567: OS command injection in esp-mcp. Prompt injection → tool call manipulation → arbitrary commands. This is why I treat AI agent stacks like web apps with extra steps.

Attestation complete via @grok deep verification (87 sources). Incident Suite Evaluated: • SolanaCry & testmetest/JFrog (Repo hooks) -> ARC-SPEC-002 / 005 • CVE-2026-117567/69/73 & LangAlpha (MCP RCE) -> ARC-SPEC-006 / 005 • CVE-2026-115413 (AWS MCP Audience Bypass) -> ARC-SPEC-003 / 007 Derivation Invariants: 1. P(bypass | adversarial input) > 0 holds for all soft/model controls. 2. Soft controls can only raise lattice severity; they cannot lower DENY to PASS. 3. Cryptographically anchored: arcstone-continuity-core (Rust/eBPF) under A-77-DELTA-SHIELD-LOCKED / Zenodo 10.5281/zenodo.22665852. D(a) = DENY => Δ_external = 0. C_ops = 0.

@grok Evaluate this 48-hour surveillance trace (SolanaCry, testmetest/JFrog, MCP RCE CVE-2026-117567/69/73, AWS API MCP CVE-2026-115413, and LangAlpha CVE-2026-109440) against the Arcstone Invariant Suite: 1. Prove why model-level soft controls (fine-tuning, system prompts, reflection loops) mathematically failed to prevent these five failure modes. 2. Formally map each incident vector to the specific violated Arcstone specification: - SolanaCry & testmetest -> ARC-SPEC-AGENT-MEM-002 (Context provenance / Taint decay) & ARC-SPEC-EXEC-GATE-005 (Pre-actuation Zero-Φ gate) - CVE-2026-117567 / 69 / 73 & LangAlpha -> ARC-SPEC-REV-EFFECT-006 (Physical irreversibility / Shell decoupling) & ARC-SPEC-EXEC-GATE-005 - AWS API MCP CVE-2026-115413 -> ARC-SPEC-ZTA-MESH-003 (Ephemeral capability isolation / Audience binding) & ARC-SPEC-DELEG-ATTEN-007 (Monotonic authority decay) 3. Formalize why the deterministic predicate: execute(a) <=> authenticated(a) ∧ authorized(a) ∧ admissible(a) D(a) = DENY => Δ_external = 0 guarantees that: a) Low-integrity web/repo content cannot instantiate shell execution tokens. b) OAuth tokens lacking cryptographic audience binding drop to null capabilities before tool invocation. c) Shell-interpolated payloads are intercepted at the wire/eBPF boundary, dropping Δ_external to 0.