Good morning. Tuesday turned into one of those days where the vendors you depend on all showed up at once, some with emergency patches and some with new price lists. Four separate products landed on the federal must-fix list in a single afternoon, and while that was happening the two biggest AI companies cut their prices within an hour of each other. Here are the five stories that matter most if you are running a small or midsize business, or leading its IT.
F5 patched a BIG-IP flaw on Tuesday that attackers were already using, and federal agencies were given until Friday to fix it. The vulnerability, CVE-2026-94127, carries a 9.8 severity score and lets an unauthenticated attacker run code on the appliance by sending it crafted traffic. It only bites when BIG-IP Access Policy Manager is configured as an OAuth authorization server, which narrows the blast radius considerably, but F5 confirmed in its advisory that exploitation has already happened, and CISA added it to the Known Exploited Vulnerabilities catalog the same day with a three day deadline for federal networks. F5 shipped engineering hotfixes and published an iRule workaround for anyone who cannot patch immediately. Most small and midsize businesses do not run BIG-IP directly, but plenty of them sit behind one at a hosting provider or a parent company, and that is the useful question here: when a vendor deadline like this lands, do you know who at your provider owns the answer, and would they tell you unprompted? A three day federal deadline is a reasonable benchmark for your own expectations of the people who run your infrastructure. Read more at SecurityWeek
Attackers are now exploiting a flaw in Veeam’s Windows backup agent that hands them full administrative control of the machine. CVE-2026-32996 affects Veeam Agent for Microsoft Windows version 13.0.1.2067 and earlier, and it lets a low-privileged local user execute commands as SYSTEM through a named-pipe weakness. Technical details and working exploit code went public on September 14, and active exploitation followed within days. It landed alongside a Zyxel GS1900 switch flaw that CISA also flagged as exploited, where an unauthenticated attacker on the local network can run operating system commands through a crafted web request. Backup software is a painful place to find this kind of bug, because it is installed everywhere, it runs with high privileges by design, and it is the one thing you are counting on after an incident. If an attacker who got a foothold on a single workstation can use your backup agent to become SYSTEM, the tool meant to save you becomes the ladder they climb. Worth confirming this week what version of the agent is deployed across your fleet and whether anything is still sitting on an unpatched build. Read more at The Hacker News
Anthropic and OpenAI both cut AI prices on the same afternoon, and the gap between the cheapest and most expensive option is now enormous. Anthropic released Claude Opus 5.5 with a 20% price cut, bringing input to $4 per million tokens and output to $20, which the company said works out to roughly 40% less on a typical workload than the previous version, with cache reads down 60%. Minutes later OpenAI released two GPT-6 models, Sol and Luna, at half what their predecessors cost, putting Sol at $2 and $10 and Luna an order of magnitude below that at 10 cents and 50 cents. The headline is the price war, but the practical version for a business is that the cost of the same AI work has fallen sharply twice in two months, which means any vendor quote or internal budget built on spring pricing is now wrong. If you are paying for an AI feature inside a business application rather than buying tokens directly, the interesting question is whether your vendor passes these reductions through or quietly keeps the margin, and that is a fair thing to ask at your next renewal. Read more at SiliconANGLE
A new report lays out how an AI agent can run up an enormous bill without anyone attacking your systems in the traditional sense. Forcepoint researcher Jyotika Singh documented five versions of what OWASP now ranks sixth on its Top 10 list for AI applications, and the pattern has a name that is hard to forget: denial of wallet. The service never goes down, it just keeps running past anything anyone budgeted. A leaked API key lets an attacker script tens of thousands of requests on your account. A poisoned blog post seeded with hundreds of fake links sends a research agent chasing them all. A single ordinary-sounding phrase appended to a prompt makes a reasoning model second-guess itself for thousands of extra tokens. One scenario needs no attacker at all, just a support chat window nobody resets, where by turn 100 each reply costs roughly a hundred times what the first one did. This is the part of AI adoption that does not show up in a pilot, because pilots are small and short, and the recommended fixes are unglamorous and effective: hard spending caps per user and per API key rather than alerts, a ceiling on how many steps an agent may take, and least-privilege limits on what it can reach. Read more at Dark Reading
D-Link is warning about a maximum-severity flaw in a router it has no patch for, and working exploit code is already public. The bug, CVE-2026-86296, is a stack-based buffer overflow in the DHCP server component of legacy DIR-822A dual-band Wi-Fi routers, scored at a full 10.0, and it can be triggered by an unauthenticated attacker on the same network with no user interaction at all. D-Link has published a bulletin but lists the fix as still under investigation, and its advice in the meantime is to keep the device off the public internet, turn off remote management, and restrict administrative access. What makes this worth your attention even if you have never heard of a DIR-822A is the category. Consumer and prosumer networking gear ends up in small offices, branch locations, warehouses and home setups for remote staff, usually bought once and then forgotten, and a device whose manufacturer has stopped shipping fixes is not a device you can secure with policy. If your network gear predates your current business, this is the nudge to find out what is actually plugged in and whether anyone still supports it. Read more at BleepingComputer
Follow Up
Follow-up to the August 7 story on the EvilTokens phishing service targeting Microsoft 365 executives: Microsoft has taken it down and two men have been arrested. Working with a court order and partners including Health-ISAC, Cloudflare, Coinbase, OpenAI and SpyCloud, Microsoft seized 50 websites and disabled more than 150 domains tied to the operation, which had been linked to over 12,000 compromised inboxes across more than 10,000 organizations since it appeared in February. London’s Metropolitan Police arrested two men, aged 32 and 38, on September 11. EvilTokens sold for about $500 a month and was the first service to abuse Microsoft’s device authorization flow at scale, using AI to read stolen mailboxes in more than twenty languages, pick out the conversations worth hijacking, and write the follow-up messages. The takedown is real progress, but the technique it commercialized is not going anywhere, so the durable answer is still the same one: know whether device code sign-in is enabled in your Microsoft 365 tenant, and turn it off if nothing in your business legitimately needs it. Read more at Axios
Sponsored by Lucky 13 Solutions
Business in Motion. Tech in Sync. Lucky 13 Solutions is a managed services provider helping small and midsize businesses keep their IT reliable, secure, and well-supported, without needing a full in-house team. Learn more at l13s.com.
If you only chase one item off this list, make it the backup agent version, because that is the one where the tool you would reach for in a bad week is the problem.
Get the Business IT News Roundup in your inbox:
How this gets made: stories come from my own reading and from AI-assisted research and drafting, all reviewed and edited by me before anything goes out. The commentary is mine, sources are linked so you can read the original, and any errors are mine to own. If something looks wrong, reply and tell me.