October 15, 2026 — Cisco's Malware Scanner Is the Way In

Host0:30

Here's a sentence nobody wants to read about their mail gateway. A crafted email attachment can lead to commands running as root. That's Cisco Secure Email Gateway, and it's being exploited right now. The flaw is CVE twenty twenty-six, seventy-six five nineteen. Cisco rates it nine point eight out of ten, and a Cisco spokesperson told The Register the company is aware of active exploitation. Here's the mechanism, per BleepingComputer. The bug sits in the Advanced Malware Protection feature, and it's a relative path traversal. Feed the scanner an attachment built to exploit that flaw, and you can get arbitrary file writes and root commands. Yes, the feature that inspects suspicious attachments is the one the suspicious attachment attacks. I'll let you sit with that. Not every gateway is exposed. You're affected when File Reputation is enabled and assigned to a mail policy, and Content Scanner Tools is earlier than twenty-five point three point zero point five five two five. You can check from the command line. Run contentscannerstatus for the version, then look at whether File Reputation sits on a mail policy. There's no workaround. The Register says Cisco recommends AsyncOS sixteen point five point zero, seven eighty, for anyone on a version earlier than sixteen point five. So the choice is patch, or hope. Now the part Cisco isn't saying. It told The Register that exploitation hit a limited number of customers. It declined to say how many appliances were compromised, or what attackers did once they were in. Limited is the vendor's word, and without a count, nobody can check it. Reports also say attackers with root may remove evidence. So a clean-looking local log proves less than you'd like. And don't mix this up with September. That Cisco advisory, CVE twenty twenty-six, seventy-six four sixty-one, was a different bug. It was an SQL injection in email parsing, also nine point eight, with advice to review mail logs for suspicious SQL. Different flaw, different trigger. Still unknown is when the attacks began, and what persistence or lateral movement followed. If you write code that writes files from untrusted input, this is your reminder to test those paths. But if you run one of these gateways, patching is today's job, not Friday's.

3:08

Cisco wouldn't give a number. SpaceXSI gave plenty, though most of them are its own. Start with the one that isn't a claim. Eight cents a minute. That's the listed price of Grok Voice three point zero, announced Wednesday, and it's the same as the version it replaces. That works out to four dollars eighty an hour, plus four tenths of a cent per text input, per the pricing page. It's a speech-to-speech model, so audio goes in and audio comes out. You reach it through the Voice Agent API over a WebSocket. You stream audio and text both ways, and you configure server-side voice activity detection, codecs, voice, instructions and tools. The docs quick-start connects to the realtime endpoint with the grok voice latest alias and an API key, or an ephemeral token from a browser. What's new for callers is mostly nothing, which is the point. The docs say the endpoint, session parameters, events and price are unchanged. The latest alias now resolves to three point zero, so if you use it, you were switched automatically. If you pinned grok voice two point zero, that ID is deprecated and you need to change it. Now the benchmarks. SpaceXSI reports an Artificial Analysis index score of eighty-six point nine. That compares with eighty-one point three for Grok Voice two point zero, eighty-three point two for GPT-Live one, and eighty-two point six for Gemini three point eight Live. The index blends four equal parts: speech reasoning, agentic performance, arena preference and task success. One piece is Tau-Voice. It runs simulated customer-service calls with tool access, then checks whether the final database state matches the expected one. SpaceXSI reports seventy-one point six percent there, and ninety-seven percent on Full Duplex Bench. But these are SpaceXSI's presentation. Artificial Analysis's page explains the methods, yet the Grok row didn't show up in what I read. The under-one-second latency claim is also the company's own, with no test conditions given. Same for the Starlink story. SpaceXSI says its call centers handle twenty thousand calls a day with Grok Voice, and about seventy percent resolve without a human. The pages read don't independently verify that. What's missing from the docs: a context window, rate limits, known weaknesses. Availability on Bedrock, Vertex or OpenRouter is unknown, and no open weights are listed. Verdict. If you're prototyping a voice agent, try it today, and measure latency and task success on your own calls, interruptions included. If you're in production, wait. An aggregate index won't tell you how it handles your customers.

6:08

That one lives behind an API and bills by the minute. This next one runs on your laptop, with no account. Debian's package for it is named dig-up with a hyphen, because Debian already has an unrelated package called digup. The project itself is DigUp. It's an on-device file search app for Mac, Windows and Linux, MIT licensed, built on Google DeepMind's EmbeddingGemma two. You point it at folders. It indexes images, PDFs, documents, audio, video, and optionally source code. Then you search by meaning, or by exact word. Here's how it works. File readers convert content, and EmbeddingGemma two runs through llama.cpp in a helper process. That uses Metal on Mac and Vulkan on Windows and Linux, where available. SQLite stores the vectors and a keyword index. Your query gets embedded, compared against the vectors, and matched against exact terms. The media handling is the fun part. Audio is processed in overlapping thirty-second windows. Video is indexed from frames sampled every few seconds, plus its soundtrack. Code gets its own index, and it's off by default. The appeal is the privacy pitch. The README says indexed files aren't uploaded, and that no account or telemetry is required. The first run downloads eight hundred sixty-five megabytes of model, and the update check can be turned off. Here's the honest limit. Those privacy claims are the project's own. I saw no independent audit and no benchmark. How search quality and indexing speed hold up on a big, mixed archive is untested. To try it on an Apple Silicon Mac, run brew install with the cask flag, and the cask is digup from the A Rahim three tap. There's also a DMG, plus Windows and Linux installers. Start with one folder, not your whole drive.

8:13

DigUp runs its model on your own machine. So does Ollama, and its newest build is about replaying tool calls. Claude Code two point one point three hundred. What you'll notice is that your prompt box can now show text that never gets sent. The new virtual text option in a mod's prompt decorations draws text in the box without entering the draft or the submitted prompt. There's also a hook that fires when the box changes without typing, like recalling history, clearing, undoing or a plugin fill. And trace spans now carry OpenTelemetry attributes for provider, conversation, tool name and error type. Upgrade if you write mods or ship traces to a collector. The release notes include no runnable example, so expect some trial and error. And the release fixes typed words like up, down and escape being read as keypresses when input arrived in a burst. So typing the word escape was, until now, a hazard. Ollama zero point forty point five, release candidate zero. What now works is replaying Responses history that contains custom tool calls through Ollama's OpenAI-compatible interface. The release notes give no code example, so you'll be testing your own history shapes. It's a release candidate, so check whether a stable build has replaced it before you adopt it. Unless you replay those histories, you can wait. The same candidate also includes a change to clarify cloud payment errors, which is a changelog line I'd like to see more of.

9:51

That's the careful end of the day. The rest of Wednesday was less careful, so here's Fast Boot. Three stories, and in one of them a bot argues its own human down to about sixty thousand dollars. Chrome exploit. XBOW says four researchers and GPT five point six Sol built a Chrome exploit in six days, earning a two hundred fifty thousand dollar bounty. A bug Google already knew disqualified chain one, so they used backups, more disaster planning than most startups have. Phishing takedown. Microsoft and partners disrupted three phishing kits that Microsoft ties to over one billion emails and texts since January. The kits also used fake CAPTCHA gates, so the scam started with a quick check that you were human. Bots at work. Comedian Graham Dot Zip's AI agent, Victor, interviewed with an AI recruiter and offered to take the job for about sixty thousand dollars. Then the bots booked a follow-up without him and got stuck saying goodbye, which is the most human thing either did. So a full-chain Chrome exploit took six days, and two bots still can't hang up a phone call.

11:05

Today we covered Cisco's actively exploited Secure Email Gateway flaw, Grok Voice three point zero, DigUp's local file search, the Claude Code and Ollama releases, and a Chrome exploit, a phishing takedown and two bots that couldn't say goodbye. Try this today. If you run a Cisco gateway, run contentscannerstatus, and if you don't, point DigUp at one folder and see what it digs up. Links to every source are in the show notes. Follow The Daily Reboot wherever you get your podcasts, and we'll be back tomorrow. I'm your host. Go patch something.