October 18, 2026 — A Backup App Held the Keys to 173 Tenants

Host0:42

The attacker worked through the tenants in sorted order, a few seconds apart. Huntress says that looks like a script walking a target list. That script had a very good key. Huntress reported unauthorized sign-ins in one hundred seventy three Microsoft three sixty five tenants, all using a compromised ConnectWise Cloud Backup credential to get Graph tokens. Microsoft saw broad successful authentication from one address on October fifteenth, then Graph reconnaissance and attempts to add credentials. The mechanism is the part to keep. The backup service is an Entra service principal, an application identity that each customer tenant granted permissions. Whoever holds its credential can request tokens and inherit those permissions. No user signs in, so MFA never gets a vote, BleepingComputer reported. Todyl says that access could read and write mail, OneDrive and SharePoint files, and Teams chats. Huntress's Jonathan Semon said the app's credential is in hostile hands. Now the counts, because there are three. One hundred seventy three is Huntress's tally of confirmed unauthorized sign-ins. BleepingComputer reported at least one hundred sixty two, and the sources don't reconcile that gap. Huntress later reported almost one thousand Entra tenants with sign-in activity from that address, and more than three hundred successful ones. A missing service principal blocked access in some. None of these is a count of stolen data, and how the credential was obtained is still unclear. ConnectWise disabled seven application identities at seven a.m. UTC on Saturday. It says it has no evidence stored backup archives were compromised, but it also acknowledged that backup jobs, archive access, and restores were unavailable. And disabling the apps doesn't reverse changes already made. Issued tokens may work until they expire. So check your tenant. In the Entra admin center, open service principal sign-ins and filter for the backup app's application ID, listed in the show notes. Then review credentials, permissions, and role assignments. That story is about a decision you can't take back. The next one is a decision you can test on your own prompts before you make it.

3:11

If you call DeepSeek's old Flash model names, you're already talking to a new model. DeepSeek says deepseek-v4-flash and its vision preview name now route to a snapshot released Saturday, V four point one Flash ten seventeen. It's a dated snapshot, not a new family, and Vercel calls it a minor update. DeepSeek claims stronger benchmarks, faster tasks, and shorter outputs, and calls it a Pareto improvement in performance and cost. Its release page shows a cost chart, but I found no numeric comparison in the text. For callers, the interface is familiar. You send chat completions in OpenAI or Anthropic format, using the model name deepseek-flash. The API documents tool calls and JSON output, in thinking or non-thinking mode, and it accepts images. Context is one million tokens. DeepSeek's API caps output at three hundred eighty four thousand, while OpenRouter lists a higher limit, so check your endpoint. Per million tokens at peak, input costs thirty cents on a cache miss and six tenths of a cent on a hit. Output is one dollar twenty. Off-peak is half. Peak means one to four and six to ten UTC, weekdays. The number that matters comes from Artificial Analysis, as listed on OpenRouter. It scores this snapshot thirty nine point five on its Intelligence Index. I found no same-evaluator comparison against the predecessor or the closest rival. What's new is the training. DeepSeek says reinforcement learning with self-feedback targets whole projects, and its showcase includes a Portal two recreation in Godot. As release-note stress tests go, that's a bold one. Limits. Concurrency tops out at twenty five hundred. OpenRouter says this snapshot isn't available through EU in-region routing. Weights are on Hugging Face and ModelScope, but the license is unknown from the sources. Verdict. Try it today on coding-agent work, long context, and chart questions, and measure latency, output tokens, and task success on your own prompts. Wait before moving production, because the improvement claims are DeepSeek's own. A model that finishes whole projects needs someone to test what it builds. This next repo pitches a board an agent can run a thousand times.

5:45

A fourteen point three million parameter transformer is steering fish around an aquarium on an ESP32-S3. In your browser, slower than real time, the project admits. That's one demo in ESP32sim, an open-source Rust emulator for real ESP32 firmware. It boots the chip's mask ROM, then the real second-stage bootloader, then your unmodified application. It covers the ESP32-S3, C3, and C6, across Xtensa and RISC-V. It models displays, touch, audio, camera, and Wi-Fi, with a virtual access point that can reach the real network. The browser build is WebAssembly, so demos run on your machine, not a server. First load is about eight megabytes. The command line is the interesting half. You feed it a script of hardware inputs, and it hands back console output, display PNGs, audio files, pin traces, and coverage data, deterministically. The README's pitch is no hardware, no cloud, no subscription. The honest limit. Three chips, modeled peripherals, and speed and hardware-match numbers that are the maintainers' own. Emulation can't promise your RF, electrical, or timing behavior. It's MIT licensed, and unofficial. To try it, open the project's browser demo and pick one. The README has a full command line for local runs. An emulator shows you what your code does when the inputs behave. The next story is about code that did something nobody intended, for six years.

7:31

The dangerous callback has been in Drupal since at least version eight point nine point zero, back in twenty twenty. Six years. Drupal just patched two critical remote code execution flaws, one in core's Field UI module and one in the Entity List contributed module. The fixes are Drupal eleven point four point ten, eleven point three point nineteen, and ten point six point nineteen, plus Entity List three point zero point ten. The Hacker News describes the mechanism. A table pre-render step ran whatever PHP callable sat in a row's region callback property, unchecked. The fix reads the row's region value instead. Here's the catch. Drupal rates the Field UI flaw nineteen out of twenty five and says core alone isn't directly exploitable. Contributed modules or custom code can expose a path, and one known path needs no authentication. The security team warned that exploits might be developed within hours or days. That's a risk assessment. As of October fifteenth, The Hacker News reported no public exploit code and no public attack reports. One number matters. Of more than five hundred thousand sites reporting a core version, twenty eight point five percent were on affected branches getting no fix. Those branches have no listed fix. If you can't patch, Drupal advises uninstalling Field UI. Drupal Steward blocks some known attack vectors, not all. So check your core branch, update, and bump Entity List separately. Six years for one callback to get noticed. Microsoft's next story runs on a faster clock, thirty eight days of training.

9:22

Two trillion parameters, says SiliconANGLE's headline. Two point eight trillion, says Microsoft's own post, and that's the figure to trust. MAI-Base-2 is a mixture-of-experts foundation model with eighty billion active parameters. Mixture of experts means each token wakes only a subset of the model. Microsoft uses it in every layer. It says its multi-round dropless routing keeps memory fixed under uneven load without dropping tokens, and that dropped tokens hurt performance. The training run took sixteen thousand GB200 GPUs and thirty eight days. Average model FLOPs utilization was twenty one percent. Goodput, the share of time training actually progressed, was eighty nine percent. The remaining eleven percent wasn't counted as training progress. Now the headline claim. Microsoft says its recipe was four point two to twelve point three times more FLOPs-efficient than DeepSeek V four Pro Base. That rests on Microsoft's own scaling-law estimates and held-out evaluations, including its own codebase and vendor-commissioned tests. The brief cites no outside reproduction. And this is a base model, still in post-training. Microsoft says it intends to license the weights to trusted partners through Foundry, with no date given. So treat it as context, not something to build on. Watch for outside evaluators testing that efficiency range. And if sixteen thousand GPUs isn't in your budget, the next four items cost you an update command.

11:05

uv zero point thirteen point one. This is the one that can change a result. uv pip check, and uv pip install with strict, now ignore Requires-Python upper bounds, matching what uv's resolver already did. If a CI step relied on those checks, re-test it. The release also adds CPython three point sixteen alpha one and Linux SPARC support, which is a sentence I did not expect to read today. Claude Code two point one point three oh two. A new environment variable picks the model for the client-side auto-mode classifier. It does not change server-side auto mode, so it's not a general override. Config gets an idle compaction toggle, and the Prompt is too long error when you stop a turn mid-compaction is fixed. Startup is faster with many plugins, especially on Windows. Ollama zero point forty point six. It adds GGUF Q two zero tensor and file types, and fixes orphaned runners on Windows. The release gives no supported-model list or benchmarks, so loading a model and watching memory is on you. Windows users should update. It's two kinds of orphan prevention in one patch, a low-bit format and a stray process. React nineteen point three point one. Released October fourteenth, with one listed change, improve transport protocol. That's the entire entry, and it doesn't say what changed. If you run React Server Components, update and run your integration tests.

12:41

Fast Boot. Three items, and one of them deletes itself. Linux checksum bug. Pwn.ai says a deferred write changed an IPv4 header from twenty bytes to sixty and exposed up to thirty-two bytes per active case, in a lab it owns. The hypervisor got off easy, because the bug lived in Linux networking. Malicious imgdecode package. imgdecode is a nine-line npm package with nine hundred seventy nine downloads that hides a Windows loader for encrypted shellcode. Researchers infer the operator can disarm it by serving a payload that decrypts into an image, so the kill switch is a JPEG. Self-erasing script. Simon Willison highlighted Lean Rada's trick, a one-line helper that lets an inline script replace itself with its output, like the current year. Unlike an orphaned Windows runner, it knows when to leave. So that's one script that vanishes to help you, and one package called imgdecode that wasn't helping anybody.

13:47

Today we covered a ConnectWise backup credential that reached one hundred seventy three Microsoft three sixty five tenants, DeepSeek's Flash snapshot taking over its old names, ESP32sim running firmware in a browser, Drupal's two critical flaws, and Microsoft's MAI-Base-2. Your try-it for today. Open Entra's service principal sign-ins and filter for the ConnectWise backup app's ID. Sources for every story are in the show notes. Follow The Daily Reboot wherever you get your podcasts. Thanks for listening, and see you tomorrow.