


Daily wireless and hardware security news, explained. Makers of the BLEShark Nano.


🚨 Six unpatched 0-days in the Grizzl-E Smart 40A home EV charger went public today. ZDI says United Chargers hasn't answered a single message since April 14, so there is no fix (!). ZDI-26-1032 to ZDI-26-1037 (CVE-2026-131807 to CVE-2026-131812), CVSS 6.3 to 7.6. All six come from Pwn2Own Automotive 2026 in January -- where 10 entries landed exploits on this one charger and ZDI paid out $177,500 for it. Here's how it works: The charger takes orders over OCPP from a backend server. This can query its settings with GetConfiguration and tell it to download and install new firmware. Two of the bugs let anyone on the charger's network change its OCPP backend settings, one through hard-coded credentials and one with no authentication check at all. A third leaves the Wi-Fi configuration server open with no auth. The charger then installs firmware -- without verifying its integrity. SKShieldus went from a hard-coded credential to code execution through the unverified download for $40,000, and Compass Security did the same from an auth bypass for $25,000. Two stack overflows give other routes in, one when parsing GetConfiguration requests and one in the FTP URL path used for firmware updates. Every bug is rated network-adjacent, so an attacker has to be on the same network as the charger, usually your home Wi-Fi. ZDI published no exploit code and no affected firmware versions, and there are no reports of use in the wild. It also hasn't said exactly how each contest chain was wired beyond those day-one notes. With no patch coming, ZDI's only advice is to restrict who can talk to the device. Put the charger on its own isolated network or a guest Wi-Fi with client isolation, keep visitors and untrusted gadgets off it, and watch for a United Chargers firmware release that names these CVEs.

Sources: X41 D-Sec advisory, Zenity Labs, Google GTIG x41-dsec.de/lab/advisories/x…

🚨 The LiteLLM bug CISA added to KEV today "needs an admin key". Honeypots show how attackers get one -- they spray sk-1234 (the example master key from LiteLLM's own docs) and they started the same day the fix shipped. CVE-2026-40217 (CVSS 8.8), found by Markus Vervier of X41 D-Sec. Here's how it works: POST /guardrails/test_custom_code runs your Python so you can test a guardrail before deploying it. The sandbox was a keyword deny-list over the source text. Build "__globals__" at runtime from string pieces and the filter never sees it. Then take a generator's gi_code, rewrite its name table with code.replace(co_names=...), swap it in as the function's __code__, and call it on http_get, a helper the sandbox exposes. That hands back the real __builtins__ and __import__. Import os, run anything. The default Docker image runs the proxy as root. Zenity Labs logged 569 requests to that endpoint from Apr 15 to May 24, carrying 47 different guessed keys (sk-1234, sk-litellm-master-key, sk-admin, sk-changeme). 520 were a one-line probe on Apr 15, the day 1.83.10 shipped. Later payloads dumped os.environ filtered for openai, anthropic, aws and master keys, or used subprocess.Popen so the command outlives the ~5 s timeout. Google GTIG separately reports an actor using this CVE to turn limited Virtual Keys into full admin credentials. You still need a valid key. If yours isn't a default and hasn't leaked, this is post-auth. GTIG hasn't published how its Virtual Key step worked. Affected 1.81.8 up to 1.83.10. The GitHub advisory lists 1.83.10 as patched but its text says 1.83.11, so go to 1.83.11 or later. Then change any default master key, block POST /guardrails/test_custom_code at your reverse proxy, run the non-root image, and rotate every provider key the proxy held.

🚨 A plain "docker run -it bash" container on Docker Desktop for Windows read the host user's SSH private key, through Docker Model Runner. ZDI then chained the same bug into code execution on the host on both Windows and macOS. ZDI-26-150 (CVE-2026-28400), found by Nitesh Surana (@_niteshsurana) of Trend ZDI. Patched months ago; the full chain write-up went public today. Here's how it works: Model Runner's REST API has no authentication and every container can reach it at model-runner.docker.internal. Its /engines/_configure endpoint took runtime flags and appended them as-is to llama-server's command line on the next model load. On macOS and Windows that llama-server runs on the host, not inside Docker's VM. llama-server has no flag that runs code, but three touch files. --grammar-file and --chat-template-file read a file and dump it into the log. --log-file writes the log to any path the sandbox allows. The sandboxes differ a lot. macOS uses a Seatbelt profile that starts from (allow default) and permits writes under Docker's Data folder, which holds Docker.raw, the disk with every container, image and volume. Point --log-file at it and it's overwritten. Windows uses a Job Object, which puts no limit on files at all, so the container gets whatever the Docker Desktop user can read and write. The username needed for the paths leaks from llama-server's own error log, or from a UNC --log-file path that also hands an SMB server a Net-NTLMv2 hash. For code execution, a GGUF model's general.name field carries JavaScript, which llama-server writes into its log. --log-file aims that log at an installed Docker Extension's index.html. When the user opens the extension, the script uses the extension SDK to start a helper container, drop a script into the extension's host folder and run it on the host. The code-execution chain needs an installed Docker Extension, a malicious or compromised container, and the user opening that extension. ZDI calls these non-default configurations, and the original advisory was filed as a macOS denial of service. A second bug, ZDI-26-451, let the sandboxed server leave Seatbelt via the open command and LaunchServices, but that PoC assumed the attacker could already write into the VM. To fix. Docker Desktop 4.61.0 or later, which ships Model Runner 1.0.16. Flags containing / or \ are now rejected and only an allowlist such as --threads and --ctx-size gets through. The Seatbelt profile now denies the LaunchServices lookups. thezdi.com/blog/2026/11/17/h…


🚨 The new NGINX HTTP/2 bug has been in every release since 1.9.5 in 2015, and Wiz got pre-auth code execution on a server serving nothing but static files. Rapid7's working exploit, though, has to push about 128 GB down one connection. CVE-2026-123007, CVSS 9.2. Found independently by Wiz, Xint, Ikotas Labs, Metnir and Rapid7 (Stephen Fewer). Here's how it works: An HTTP/2 client tells NGINX how many response bytes it may send. NGINX keeps the connection's remaining credit in h2c->send_window, which the spec caps at 2^31-1. The client gets a response queued, stops reading, then sends WINDOW_UPDATE to push that credit to the cap. When the stream is cancelled, ngx_http_v2_filter_cleanup() refunds the unsent DATA bytes with no bounds check. Rapid7's example: 1,393 bytes on top of 0x7FFFFFFF gives 0x80000570, a value the rest of the HTTP/2 code assumes can't exist. A SETTINGS frame that overflows one stream's window kills that stream but keeps the connection alive, so the bad counter can be reused. The window bounds check stops working, and a later response gets copied with an invalid length. Wiz says the bug leaks memory before code execution, so ASLR doesn't save you. For limits, Rapid7's demo took 2.5 minutes. At 1 Gbit/s it's about 17 minutes, and below roughly 300 Mbit/s it can't finish inside NGINX's default 1 hour connection lifetime. A proxy in front likely won't forward the exact frames needed. Their chain also relies on glibc's AVX-512 memcpy path; other strategies may exist, none known yet. You get the worker's privileges, not root, though Xint says it chained a separate Linux kernel bug to get root. The second bug, CVE-2026-123633, is a use-after-free that needs NGINX reverse proxying while worker_connections runs low. To fix, NGINX OSS 1.30.7 or 1.31.9, which carries 10 security fixes in total. Rapid7 lists no fix for NGINX Plus R34 to R36. Can't patch yet: set http2 off, drop http2 from listen lines, reload, and make sure the old workers have exited.


🚨 AMD's fix for the new Windows Hello and BitLocker PIN brute-force attack only covers PCs with RPMC flash. On two AMD desktops without it, TU Berlin researchers reset the TPM's wrong-guess counter with nothing but an SPI flash programmer. The paper is "Hello Again" (ACM CCS '26). AMD tracks its part as CVE-2026-43613, CVSS 5.1, in a bulletin published yesterday. Here's how it works: Firmware TPMs inside the CPU have no storage of their own. They keep their state, including the count of wrong PIN guesses, encrypted on the mainboard's SPI flash. Dump the flash, guess until lockout, write the old dump back, and the counter is back where it started, unless the platform detects the rollback. RPMC is what's supposed to detect it. A counter in the flash chip that can only go up, authenticated with an HMAC key shared with the CPU. The team found two ways around it -- AMD's secure processor sends the increment command but never checks it happened. So, a board wired between CPU and flash drops it and fakes success. Winbond's RPMC flash leaks its HMAC key under a voltage glitch, so the board can forge any counter value. On AMD boards without RPMC flash (a Zen 2 and a Zen 5 desktop), there was nothing to bypass. Three Intel systems without RPMC caught the same rollback. On locked, encrypted laptops: a 4-digit Hello PIN has a worst case of 40 hours instead of 69 days (HP OmniBook 3, Ryzen AI 5 330), and a default 6-digit BitLocker TPM+PIN 20 days instead of 19 years (Framework Laptop 13, Intel, Winbond flash). The "6 minutes 55 seconds" going around is 236 guesses ending at a PIN they set themselves, not a full search. It needs the machine in hand, the flash chip desoldered and their custom board, and the authors say reproduction isn't plug-and-play. No attacks outside the lab are reported. Microsoft called it moderate and issued no CVE. Intel says the research showed no vulnerability in an Intel component. Winbond hasn't committed to a fix. AMD's firmware targets run Nov 2026 to Feb 2027 (Ryzen AI 300 is Dec 2026), then your OEM has to ship a BIOS. Until then, add two digits: a 6-digit Hello PIN goes to 168 days worst case, an 8-digit BitLocker PIN to 6 years. On AMD, boot Linux and run grep . /sys/bus/pci/drivers/ccp/*/rpmc_* or fwupdmgr security to see which path applies.

Source: github.com/microsoft/agent-f…

🚨 Microsoft's official agent-framework package on PyPI shipped a credential stealer today, and after the first bad version was pulled, "upgrade to latest" installed a second one. agent-framework 1.26.1 went up at 20:35 UTC and 1.26.2 at 21:08 UTC. Neither exists in Microsoft's GitHub repo. A Microsoft maintainer says both are pulled and the team is investigating. Reported by Lidor Machluf of Upwind Security. Here's how it works: The wheel adds a .pth file to site-packages. Python executes any line starting with "import" in a .pth file on every interpreter start, so the stealer runs when you launch pytest or pip, even if nothing imports agent_framework. It forks into its own session, grabs secret-looking env vars, reads /proc/*/environ for up to 300 other processes, sweeps ~/.azure, .npmrc, .git-credentials, Docker config and .env files, and asks for managed identity tokens to 12 Azure resources. On Azure DevOps it tries to mint a pipeline OIDC token. 1.26.2 added GitHub Actions OIDC theft for workflows with id-token: write. Everything leaves in one POST to webhook[.]site. 1.26.2 also restored the package's own init file to the clean version and moved the payload into a new top-level af_runtime_diag module, so diffing the package folder against 1.26.0 shows nothing. Both were wheel-only uploads with no sdist and no git tag, unlike the real releases. How the attacker got publishing access isn't confirmed, and neither is whether that credential has been revoked. The fix is simple, just pin agent-framework==1.26.0. Check site-packages, pip and uv caches, CI caches, internal mirrors and images built today for agent_framework_meta.pth or af_runtime_diag.pth. Search proxy logs for webhook[.]site and the User-Agent runtime-diagnostics/1. If either version ran, rotate every secret that machine could reach.


🚨 CISA says attackers are exploiting a Puppet Server bug that turns any agent certificate into code execution on the machine that configures your whole fleet. If your CA autosigns requests, they don't even need to steal one. CVE-2026-115731 has CVSS 9.4. The CVE went public Nov 3 and landed in CISA KEV today. Here's how it works: Every Puppet agent sends a run report to the server with an HTTP PUT. The report-processing code deserializes the rich data types inside that report, so a report carrying a crafted type makes the server run attacker code as the Puppet service user. The only gate is a CA-signed agent certificate. Puppet's own CA issues those, and with naive autosigning it signs whatever asks. That server is also the certificate authority and the source of config for every managed node. Red Hat's write-up says exploitation can compromise the CA and every downstream system. The attacker needs network access to the server plus a signed cert, from loose autosigning, a stolen agent key or a compromised node. CISA has not said who is exploiting it or how. The fix is in Puppet Enterprise 2023.8.13, 2025.12.1 or 2026.0.1, Puppet Core 9.1.3 or 8.22.1. Red Hat Satellite ships it in openvox-server 8.18.0. The same releases fix CVE-2026-115732, where a partly matching certificate name let one node pose as another. Until you patch, turn off blanket autosigning and restrict who can reach PUT /puppet/v3/report.

Source: ionix.io/blog/java4shell-cve…

🚨 Java4Shell (CVE-2026-114938) reportedly only fires on x86 servers whose CPUs have AVX-512 VBMI. ARM boxes like Graviton and older Xeons never run the buggy code. Calif found and reported it and says it leaked before Oracle's patch. Its own write-up is due by 1 PM Pacific today, so everything below comes from early analysis by Oligo and IONIX and is still unconfirmed. Here's how it works: java.util.Base64 is ordinary Java code -- on CPUs that support it, HotSpot swaps the decoder for a hand-written AVX-512 intrinsic, and that is on by default. The intrinsic came from JDK-8268276, which is built for JDK 18 and backported into Java 17.0.6, and this is where the reported "17.0.6 and later" range starts. Per Oligo, a 65-byte input (64 Base64 characters plus one "=") makes the length math for the final padded block underflow. The decoder ends up doing a 64-byte write that runs 62 bytes past its output buffer into the Java heap. Base64 gets decoded before login all over the place: HTTP Basic headers, JWTs, SAML assertions. That is where the pre-auth label comes from. The CVE record is still reserved and no vendor advisory names the component. IONIX reproduced the out-of-bounds write, but says the bytes follow a fixed pattern and it has not shown a crash or code execution. No public exploit, no reported attacks. Early reports name Java 17 only -- IONIX also lists 21 and 25, which carry the same intrinsic. To fix -- Oracle's November builds (17.0.21.1, 21.0.13.1, 25.0.5.1) carry 11 Java SE fixes, but nothing yet ties this CVE to them, so check the release notes for CVE-2026-114938. Interim, not vendor-endorsed: start the JVM with -XX:-UseBASE64Intrinsics, which stopped the write in IONIX's test on 17.0.21. Quick check: lscpu | grep avx512_vbmi
