critical

CVE

CVE-2026-61536, CVE-2026-54912, CVE-2026-54914, CVE-2026-54143, CVE-2026-54141, CVE-2026-54146, CVE-2026-54145, CVE-2026-54142, CVE-2026-54140, CVE-2026-49393, CVE-2026-97693, CVE-2026-123018, CVE-2026-82033

CWE

CWE-94, CWE-787, CWE-125, CWE-306, CWE-269, CWE-214, CWE-470

Affected Surface

  • PyPI `banks` (`github.com/masci/banks`, LlamaIndex prompt template library) versions `<= 2.4.2` (`CVE-2026-61536` / `GHSA-64vx-6h2c-rjh7`, CVSS 7.5 High, fixed in `2.4.3`)
  • Bytecode Alliance WebAssembly Micro Runtime (`wasm-micro-runtime/wasm-micro-runtime`, `iwasm`) versions `< 2.4.5` (`CVE-2026-54912` CVSS 8.8 High and `CVE-2026-54914` CVSS 8.3 High, fixed in `WAMR-2.4.5`)
  • Rust / Tauri AI agent harness `openhuman` (`github.com/tinyhumansai/openhuman`) versions `< 0.56.0` (`CVE-2026-49393`), `< 0.57.5` (`CVE-2026-54145`, `CVE-2026-54146`), and `< 0.57.13` (`CVE-2026-54140`, `CVE-2026-54141`, `CVE-2026-54142`, `CVE-2026-54143`, up to CVSS 9.8 Critical)
  • Python AI agent platforms `Yuxi` (`github.com/xerrors/Yuxi` `>= 0.6.0.beta1, < 0.7.1`, `CVE-2026-97693` / `GHSA-qpm5-6332-9445`, CVSS 9.9 Critical), `DB-GPT` (`github.com/eosphoros-ai/DB-GPT` `<= 0.8.2`, `CVE-2026-123018`, CVSS 8.3 High), and `DeepSeek Harness` (`github.com/deepseek-ai/deepseek-harness` `< 0.1.2-alpha.2`, `CVE-2026-82033`, CVSS 8.4 High)

TL;DR

Between 4 November (19:17 UTC) and 5 November 2026 (18:17 UTC), NVD and the GitHub Advisory Database published 13 high- and critical-severity vulnerabilities across six AI agent runtimes, prompt-templating packages, and WebAssembly engines:

  1. LlamaIndex banks <= 2.4.2 CompletionExtension importlib RCE (CVE-2026-61536 / GHSA-64vx-6h2c-rjh7, CVSS 7.5 High, fixed in 2.4.3): banks 2.4.2 switched its Jinja2 environment to SandboxedEnvironment to fix CVE-2026-44209, but left CompletionExtension._get_tool_callable() in src/banks/extensions/completion.py untouched. Because {% completion %} parses Tool JSON objects from the rendered template body and resolved tool.import_path via importlib.import_module(module_name) and getattr(mod, func_name) without an allowlist, an attacker controlling a template or tool metadata string could set "import_path": "os.system" (or "subprocess.getoutput") and execute arbitrary OS commands when the LLM emitted a matching tool_calls entry.
  2. Bytecode Alliance WebAssembly Micro Runtime (WAMR < 2.4.5, CVE-2026-54912 CVSS 8.8 High & CVE-2026-54914 CVSS 8.3 High): In core/iwasm/interpreter/wasm_loader.c, the two-pass fast interpreter loader capped constant-table deduplication scans at FAST_INTERP_CONST_DEDUP_SCAN_MAX (32), causing Pass 1 and Pass 2 to disagree on whether a constant was already deduplicated once intervening constants pushed the table past 32 entries and overflowing the heap-allocated constant table (CVE-2026-54912, public PoC gist shumbo/b710c35b005ed871447fce385d1ab760). In core/iwasm/libraries/libc-wasi/libc_wasi_wrapper.c, wasi_poll_oneoff() checked validate_app_addr(out_offset, sizeof(wasi_event_t)) for a single event struct instead of nsubscriptions * sizeof(wasi_event_t) (CVE-2026-54914).
  3. OpenHuman AI agent harness (< 0.56.0 / < 0.57.5 / < 0.57.13, 7 CVEs up to CVSS 9.8 Critical): Prior to 0.56.0, the Rust core server required bearer auth on HTTP POST /rpc but exposed the identical JSON-RPC methods unauthenticated over Socket.IO rpc:request (CVE-2026-49393, CVSS 8.4). Prior to 0.57.5, the macOS Tauri desktop binary spawned openhuman-core with --token <OPENHUMAN_CORE_TOKEN> in argv, exposing the root RPC secret to any local process via ps / sysctl (CVE-2026-54146, CVSS 8.8), and used the static bearer token as the approval-gate session ID (CVE-2026-54145, CVSS 8.8). Prior to 0.57.13, external chat channels (such as Yuanbao DMs) could create persistent scheduled cron jobs that re-ran with full host tool privileges (CVE-2026-54143 CVSS 9.8 Critical and CVE-2026-54141 CVSS 9.0 Critical).
  4. Yuxi (< 0.7.1, CVE-2026-97693 CVSS 9.9 Critical), DB-GPT (<= 0.8.2, CVE-2026-123018 CVSS 8.3 High), and DeepSeek Harness (< 0.1.2-alpha.2, CVE-2026-82033 CVSS 8.4 High): Yuxi placed agent code-execution containers on the same Docker network as internal Postgres, MinIO, and sandbox-provisioner services without authentication (CVE-2026-97693); DB-GPT’s execute_code() in code_utils.py checked if not use_docker or in_docker_container: (where in_docker_container = os.path.exists("/.dockerenv")), silently bypassing its own Docker sandbox whenever DB-GPT itself ran inside a container (CVE-2026-123018); and DeepSeek Harness exposed its local HTTP/WebSocket host RPC API without a bearer token (CVE-2026-82033).

What to do now: Upgrade banks to >= 2.4.3 (2.5.1 latest), WAMR (iwasm) to >= 2.4.5, OpenHuman to >= 0.57.13, Yuxi to >= 0.7.1, and DeepSeek Harness to >= 0.1.2-alpha.2. On DB-GPT (<= 0.8.2), do not expose port 5670 to untrusted callers and disable CodeAction native subprocess.run fallback inside containers.

Affected packages and projects

EcosystemPackage / ProjectCVE / GHSASeverityAffected versionsFixed versionVulnerable component
PyPIbanks (masci/banks)CVE-2026-61536 (GHSA-64vx-6h2c-rjh7)CVSS 7.5 High<= 2.4.22.4.3 (2.5.1 latest)src/banks/extensions/completion.py (_get_tool_callable)
C / WASMwasm-micro-runtime (WAMR)CVE-2026-54912, CVE-2026-54914CVSS 8.8 / 8.3 High< 2.4.52.4.5core/iwasm/interpreter/wasm_loader.c, libc_wasi_wrapper.c (wasi_poll_oneoff)
Rust / Tauriopenhuman (tinyhumansai/openhuman)CVE-2026-54143, CVE-2026-54141, CVE-2026-54142, CVE-2026-54140CVSS 9.8 / 9.0 / 8.7 / 8.3< 0.57.130.57.13src/openhuman/agent/turn_origin.rs, src/openhuman/cron/scheduler.rs, mcp prompt assembly
Rust / TauriopenhumanCVE-2026-54146, CVE-2026-54145, CVE-2026-49393CVSS 8.8 / 8.8 / 8.4 High< 0.57.5 (54146, 54145); < 0.56.0 (49393)0.57.5, 0.56.0core_process.rs (OPENHUMAN_CORE_TOKEN argv), approval/gate.rs, src/core/socketio.rs
Python / DockerYuxi (xerrors/Yuxi)CVE-2026-97693 (GHSA-qpm5-6332-9445)CVSS 9.9 Critical>= 0.6.0.beta1, < 0.7.10.7.1docker/sandbox_provisioner/app.py, docker-compose.yml sandbox network isolation
PythonDB-GPT (eosphoros-ai/DB-GPT)CVE-2026-123018CVSS 8.3 High (9.8 CNA)<= 0.8.2Unpatched upstream (mitigate via network auth & sandbox config)packages/dbgpt-core/src/dbgpt/util/code_utils.py (execute_code)
Rust / Webdeepseek-harnessCVE-2026-82033CVSS 8.4 High< 0.1.2-alpha.20.1.2-alpha.2Web profile HTTP and WebSocket Host RPC API (commit 3e24087bfaea)

Technical analysis

1. CVE-2026-61536: Why Jinja2 SandboxedEnvironment did not stop banks {% completion %} RCE

In banks 2.4.2, the maintainers fixed a classic Jinja2 SSTI (CVE-2026-44209 / GHSA-gphh-9q3h-jgpp) by switching src/banks/env.py from jinja2.Environment to jinja2.sandbox.SandboxedEnvironment. However, banks also provides a {% completion %} block tag (src/banks/extensions/completion.py) that lets templates turn Python functions into LLM tool schemas using the | tool filter.

Prior to 2.4.3 (commit e443f112b80a, PR #75), the | tool filter serialized a Tool Pydantic model containing "import_path": f"{func.__module__}.{func.__qualname__}" as JSON inside the rendered block. When the LLM responded with a tool_calls entry, CompletionExtension._get_tool_callable() deserialized the Tool JSON from the rendered text and imported whatever dotted path was written in tool.import_path:

--- a/src/banks/extensions/completion.py
+++ b/src/banks/extensions/completion.py
@@ -41,6 +40,14 @@ class CompletionExtension(Extension):
     tags = {"completion"}  # noqa
 
+    # Maps tool function names to the actual callables registered via the `tool`
+    # filter. Using an explicit registry avoids calling `importlib.import_module`
+    # on attacker-controlled strings embedded in template text.
+    _callable_registry: ClassVar[dict[str, Callable[..., Any]]] = {}
@@ -143,12 +150,11 @@ class CompletionExtension(Extension):
     def _get_tool_callable(self, import_path: str) -> Callable:
-        items = import_path.split(".")
-        module_name = ".".join(items[:-1])
-        func_name = items[-1]
-        mod = importlib.import_module(module_name)
-        return cast(Callable, getattr(mod, func_name))
+        try:
+            return self._callable_registry[import_path]
+        except KeyError as exc:
+            msg = f"Tool '{import_path}' is not registered. Only callables passed through the `tool` filter may be invoked."
+            raise ValueError(msg) from exc

Because {% completion %} accepted any JSON block matching the Tool schema inside the rendered template body, an untrusted template did not need Jinja2 attribute traversal at all: embedding a raw JSON Tool definition with "import_path": "os.system" caused _get_tool_callable("os.system") to return os.system and pass the LLM’s tool_call.function.arguments directly to it.

2. CVE-2026-54912 and CVE-2026-54914: WAMR 2.4.5 fast-interpreter constant table overflow and wasi_poll_oneoff out-of-bounds write

In Bytecode Alliance’s WebAssembly Micro Runtime (WAMR < 2.4.5), when WASM_ENABLE_FAST_INTERP != 0, wasm_loader.c runs a two-pass loader: Pass 1 counts unique constants (const_cell_num) to allocate a single contiguous constant buffer, and Pass 2 populates the buffer. To avoid $O(n^2)$ scans on modules with thousands of constants, WAMR introduced FAST_INTERP_CONST_DEDUP_SCAN_MAX (32), scanning for duplicates only while const_count < 32.

However, Pass 1 incremented const_count for all constants as it scanned instructions, whereas Pass 2 emitted function parameter/local zero-initializers and branch constants in a slightly different order. Once a function crossed 32 constants, a constant that was deduplicated in Pass 1 (while const_count < 32) could be evaluated in Pass 2 after const_count >= 32, skipping deduplication, allocating a new slot in Pass 2 that Pass 1 never budgeted for, and overflowing the heap buffer (CVE-2026-54912, fixed in commit b1353b5e8983).

In the same WAMR-2.4.5 release, commit b49d10eac417 fixes CVE-2026-54914 in wasi_poll_oneoff() across both libc-wasi and libc-uvwasi:

--- a/core/iwasm/libraries/libc-wasi/libc_wasi_wrapper.c
+++ b/core/iwasm/libraries/libc-wasi/libc_wasi_wrapper.c
@@ -887,9 +887,12 @@ wasi_poll_oneoff(wasm_exec_env_t exec_env, const wasi_subscription_t *in,
     wasm_module_inst_t module_inst = get_module_inst(exec_env);
 
-    if (!validate_native_addr((void *)in, sizeof(wasi_subscription_t))
-        || !validate_native_addr(out, sizeof(wasi_event_t))
+    if (((uint64)nsubscriptions * sizeof(wasi_subscription_t) > UINT32_MAX)
+        || !validate_native_addr((void *)in,
+                                 (uint32)nsubscriptions * (uint32)sizeof(wasi_subscription_t))
+        || !validate_native_addr(out,
+                                 (uint32)nsubscriptions * (uint32)sizeof(wasi_event_t))
         || !validate_native_addr(nevents, sizeof(uint32)))
         return (wasi_errno_t)-1;

Previously, a WASI module could place in or out 1 struct before the end of linear memory and pass nsubscriptions > 1, reading or writing host heap memory immediately after the WASM linear memory mapping.

3. CVE-2026-123018 (DB-GPT), CVE-2026-97693 (Yuxi), and CVE-2026-54143 (OpenHuman): AI agent sandbox and channel escapes

In DB-GPT (<= 0.8.2), CodeAssistantAgent extracts triple-backtick python, bash, and sh blocks from LLM output and passes them to execute_code() in packages/dbgpt-core/src/dbgpt/util/code_utils.py. To decide whether to run the code inside a disposable Docker container or directly on the host with subprocess.run, execute_code() ran:

# packages/dbgpt-core/src/dbgpt/util/code_utils.py (DB-GPT <= 0.8.2)
in_docker_container = os.path.exists("/.dockerenv")
if not use_docker or in_docker_container:
    # Runs LLM-generated bash/python directly inside the DB-GPT server process!
    result = subprocess.run([cmd, filepath], capture_output=True, text=True, ...)

Because official DB-GPT deployments run inside Docker (where /.dockerenv always exists), in_docker_container is True, short-circuiting the if condition and running every LLM-generated script directly inside the DB-GPT application container (with full access to DB-GPT’s database credentials and model API keys).

Similarly, in Yuxi (< 0.7.1, CVE-2026-97693), commit b5b5713e6e63 isolates agent sandbox containers onto a dedicated internal Docker network (sandbox_net), requires an X-Sandbox-Provisioner-Token header on the sandbox provisioner API, and strips backend database and API credentials from sandbox environments. And in OpenHuman (< 0.57.13, commit 3a7180da7765), introducing TurnOrigin prevents messages arriving over external chat channels (CVE-2026-54141, CVE-2026-54143) from invoking privileged host tools or scheduling cron jobs that later execute as a local user turn.

Detection and remediation

# 1. Check Python environments and lockfiles for banks <= 2.4.2 and dbgpt <= 0.8.2
pip show banks dbgpt 2>/dev/null || true
rg -n '(banks|dbgpt)' requirements*.txt pyproject.toml uv.lock poetry.lock Pipfile.lock 2>/dev/null

# 2. Check WAMR (iwasm), OpenHuman, Yuxi, and DeepSeek Harness versions
iwasm --version 2>/dev/null || true
rg -n 'wasm-micro-runtime|openhuman|xerrors/Yuxi|deepseek-harness' Cargo.lock go.mod docker-compose*.yml Dockerfile* 2>/dev/null
  1. Upgrade banks to >= 2.4.3 (pip install -U "banks>=2.4.3"): Verify that templates using {% completion %} register all tools through the | tool filter rather than raw JSON import_path strings.
  2. Upgrade WAMR to >= 2.4.5: Any service embedding libiwasm / vmlib to run untrusted WebAssembly modules with fast interpreter or WASI enabled must update to WAMR-2.4.5.
  3. Upgrade OpenHuman to >= 0.57.13, Yuxi to >= 0.7.1, and DeepSeek Harness to >= 0.1.2-alpha.2: On OpenHuman deployments, audit scheduled cron jobs created prior to upgrading (CVE-2026-54143). On DB-GPT deployments (<= 0.8.2), place port 5670 behind authentication and patch or disable CodeAction so in_docker_container does not fall back to host subprocess.run.

Timeline

  • 2026-06-18 to 2026-08-25: Maintainers commit fixes in masci/banks (e443f112b80a, v2.4.3), wasm-micro-runtime (b1353b5e8983 and b49d10eac417, WAMR-2.4.5), tinyhumansai/openhuman (b1bbc53fceed, 3a7180da7765, v0.56.0 / v0.57.5 / v0.57.13), xerrors/Yuxi (b5b5713e6e63, v0.7.1), and deepseek-ai/deepseek-harness (3e24087bfaea, dsh-v0.1.2-alpha.2).
  • 2026-11-04 (19:17 to 22:18 UTC): NVD publishes CVE-2026-82033 (DeepSeek Harness) and CVE-2026-97693 (Yuxi).
  • 2026-11-05 (12:17 to 18:17 UTC): NVD and the GitHub Advisory Database publish CVE-2026-123018 (DB-GPT), CVE-2026-49393 and CVE-2026-54140..54146 (OpenHuman), CVE-2026-61536 / GHSA-64vx-6h2c-rjh7 (banks), and CVE-2026-54912 / CVE-2026-54914 (WAMR).

From research to remediation

Check whether this pattern exists in your codebase

Turn this research into a remediation workflow. Scan dependencies and package manifests for similar supply-chain risk, then prioritize fixes with reachability context.

References