AI and IT News Recap: July 27, 2026: Anthropic's Claude Opus 5 Undercuts the Frontier at Half the Price, an Unpatched Fastjson RCE Is Under Attack, and Cl0p Squeezes Engineering Firms

By Noah Smith, Owner & Consultant, KeyChange Technologies · July 27, 2026

Welcome to your AI and IT news recap for Monday, July 27, 2026, covering the weekend of Friday July 24 through Sunday July 26. The headline is a big one: Anthropic shipped a new flagship model that chases the frontier at half the price. Underneath it, the security news is unusually sharp, led by a critical bug with no patch in sight. Missed Friday? Catch the July 24 recap.

📌 The AI and IT news at a glance

• 🔝 Anthropic launched Claude Opus 5. It reaches roughly Fable 5-level intelligence at half the price, adds an effort toggle to trade cost for capability, and became the default model on Claude Max.
• 🛡️ A Fastjson bug is under active attack with no patch. CVE-2026-16723 (CVSS 9.0) lets a single JSON request run code on unpatched Java apps, and the 1.x branch is end-of-life, so no fix is coming.
• 🛡️ Cl0p is squeezing engineering firms through PTC Windchill. The ransomware crew is exploiting CVE-2026-12569 to steal product-design data and has started emailing victims.
• 🛡️ Malvertising now builds the malware inside your browser. A campaign called SourTrade impersonates Solana, Luno, and TradingView and assembles its payload in memory to dodge scanners.
• 🛡️ OnTrac is notifying customers of a data breach. The last-mile delivery firm says a network intrusion exposed personal data for around 40,000 people.
• 🤖 The open-weights fight went public. Nvidia's letter urging Washington not to restrict downloadable models doubled to 50 signatories in a day, with Anthropic and Amazon notably staying out.

🔝 Today's top story: Anthropic's Claude Opus 5 chases the frontier at half the price

Anthropic released Claude Opus 5 on Friday, and it reshapes the price-to-capability math that has defined the frontier all summer. The company says Opus 5 lands at roughly the intelligence of its top-tier Fable 5 model while costing half as much to run: $5 per million input tokens and 5 per million output tokens, the same rate as the older Opus 4.8 and half of Fable 5's input price. On the benchmarks Anthropic published, Opus 5 more than doubles Opus 4.8's score on the Frontier-Bench evaluation, lands within about half a percent of Fable 5's peak on the CursorBench coding test, and posts large gains on agentic tasks like OSWorld. It ships with a 1M-token context window and is now the default model on Claude Max and the strongest option on Claude Pro.

The more interesting design choice is a low/medium/high effort toggle that lets you dial how hard the model thinks on a per-request basis, trading intelligence for speed and token savings when a task does not need the full engine. Opus 5 is available in the API as claude-opus-5, and it has already been wired into Claude Code as the new default Opus model, so developers using the coding agent get it without changing anything. For a market that spent June and July absorbing expensive frontier models from every lab, a near-frontier model at half the cost is the kind of move that pulls real budget.

In short: Anthropic launched Claude Opus 5, a near-frontier model priced at half of Fable 5 ($5/5 per million tokens) with an adjustable effort toggle, now the default on Claude Max and in Claude Code.

What it means for your business: If your team already pays for frontier-tier AI, Opus 5 is a straightforward way to cut the bill or do more work for the same spend, especially for coding and knowledge work. The effort toggle also means you can run cheap, fast passes for routine tasks and reserve maximum reasoning for the hard ones, which is a practical lever for controlling AI costs.

My take: The story of 2026 is that frontier-level capability keeps getting cheaper faster than anyone budgeted for, and Opus 5 is the clearest example this month. The effort toggle is the part I would actually use day to day: most business tasks do not need a model thinking at full tilt, and paying for that by default is how AI bills quietly balloon. If you standardized on a pricier model earlier this summer, this is worth a re-evaluation.

Source: TechCrunch

🛡️ IT and security

A Fastjson bug is under active attack, and there is no patch coming

Security teams got the worst kind of vulnerability this weekend: a critical, actively exploited flaw with no fix available. CVE-2026-16723 is a remote code execution bug in Fastjson 1.x, the widely used Java JSON library from Alibaba, rated CVSS 9.0 and affecting versions 1.2.68 through 1.2.83. A malicious JSON request sent to a vulnerable application can execute code with the same privileges as the Java process, no authentication required. What makes this one especially dangerous is that it works under default settings, whereas older Fastjson bugs only triggered when the risky AutoType feature was switched on, so many previously hardened deployments are exposed again.

The catch is that Fastjson 1.x is no longer maintained, so there is no patched 1.x release and none is coming. Threat intelligence firm ThreatBook reported in-the-wild exploitation after adding detection on July 22, and Imperva has documented attacks against financial services, healthcare, and retail organizations, primarily in the United States. The recommended response is to enable Fastjson's SafeMode immediately as a stopgap and plan a migration to the actively maintained Fastjson 2.x line.

In short: CVE-2026-16723, a CVSS 9.0 unauthenticated RCE in the end-of-life Fastjson 1.x library, is being exploited in the wild with no patch available, hitting US finance and healthcare targets.

What it means for your business: If any of your software or a vendor's software runs on Java, ask whether Fastjson 1.x is in the stack, because it is buried inside countless Spring Boot applications. Enabling SafeMode buys time, but the real fix is getting off the unmaintained 1.x branch.

My take: "Critical, exploited, and unpatchable" is about as bad as a vulnerability gets, and the fact that it works with default settings means a lot of teams who thought they were safe are not. Do not wait for a patch that will never ship. Turn on SafeMode now and put the 2.x migration on the calendar this week, not this quarter.

Source: The Hacker News

Cl0p is exploiting a PTC Windchill flaw to steal engineering data

The Cl0p ransomware group, best known for mass data-theft campaigns against file-transfer tools, has moved on to product lifecycle management software. Affiliates are exploiting CVE-2026-12569, a critical deserialization flaw (CVSS 9.8) in PTC's Windchill and FlexPLM platforms that allows unauthenticated remote code execution. The bug was disclosed in mid-June and added to CISA's Known Exploited Vulnerabilities catalog on June 25, but the extortion phase is now underway: starting around July 20, victims began receiving emails with the subject line "Windchill PDMLink module serious data leak."

Windchill is used by manufacturing, automotive, aerospace, and apparel companies to manage designs and product data, which makes the stolen material unusually sensitive. Post-exploitation activity observed by researchers includes enumerating the filesystem and staging engineering and product-design files for exfiltration. Any organization running an internet-exposed Windchill or FlexPLM instance should assume it is a target and verify patch status against the June advisory.

In short: Cl0p affiliates are exploiting CVE-2026-12569, an unauthenticated RCE in PTC Windchill and FlexPLM, to steal engineering data and have begun extorting victims by email.

What it means for your business: If you use Windchill or FlexPLM, or partner with manufacturers who do, treat this as an active threat to your intellectual property, not just an IT ticket. Product designs and engineering files are exactly the kind of data that is painful to have stolen and hard to claw back.

My take: Cl0p has a reliable playbook: find a widely deployed enterprise app with a critical bug, hit everyone at once, and extort by email rather than encrypting. The lesson keeps repeating: internet-exposed enterprise applications are the front door now. If a system does not need to be reachable from the open internet, it should not be.

Source: BleepingComputer

A malvertising campaign builds its malware inside your browser

Researchers at Confiant detailed a slick malvertising operation, dubbed SourTrade, that turns the victim's own browser into a malware assembly line. The campaign runs fake webpages impersonating trading and crypto platforms including Solana, Luno, and TradingView, then uses malicious JavaScript to deliver the malware in harmless-looking pieces and instruct the browser to assemble the final executable directly in memory. Because no complete malicious file is ever downloaded in one piece, the technique slips past many scanners that watch for known-bad files.

The operation is highly targeted: a filtering system funnels genuine retail traders and crypto investors to the malicious pages while redirecting researchers, scanners, and security bots to blank pages. SourTrade has been running since late 2024 and has been localized into 25 languages across 12 countries, concentrated in Asia Pacific and Latin America. The practical defense is old-fashioned: get financial and crypto apps only from the vendor's official site, not from ads or sponsored search results, and verify the installer's digital signature before running it.

In short: A malvertising campaign called SourTrade impersonates Solana, Luno, and TradingView and has browsers assemble malware in memory to evade detection, targeting retail traders across 12 countries.

What it means for your business: Anyone on your team who handles company crypto or trading accounts is in the target profile, and this technique defeats the "just don't download sketchy files" advice because the browser does the assembly. Reinforce that software comes from official sources only, and consider blocking known malvertising infrastructure at the network level.

My take: This is a clever escalation of an old trick, and the memory-assembly step is designed specifically to beat the file-based detection most people rely on. The defense has not changed, though: ads and sponsored search results are a terrible place to get software. Train people to navigate to a vendor's site directly, every time.

Source: BleepingComputer

OnTrac notifies customers of a data breach

OnTrac, a large US last-mile parcel-delivery company, is notifying customers that attackers breached its corporate network and accessed personal data. The company says it detected the incident earlier this year and that intruders reached certain files over a short window before being cut off. Reporting puts the number of affected people at roughly 40,000, with exposed data including names and, according to some accounts, birth dates and driver's license details. OnTrac is offering affected customers 12 months of free credit monitoring and identity-protection services.

OnTrac operates across roughly 35 states and covers a large share of the US population through thousands of independent delivery contractors, which is what makes a breach at this kind of logistics middleman worth watching. Delivery and fulfillment vendors sit on a surprising amount of customer personal information, and they are an increasingly attractive target precisely because so many retailers route data through them.

In short: OnTrac is notifying about 40,000 people that a network intrusion exposed personal data, and is offering 12 months of credit monitoring.

What it means for your business: If you ship through OnTrac or a similar last-mile carrier, remember that your customers' data flows through their systems too, and a breach there can become your customer-service problem. It is a reminder to know which vendors hold your customer data and what their security posture looks like.

My take: Logistics and fulfillment vendors are the quiet soft underbelly of retail data security. They rarely make headlines until they get breached, and by then the data is already out. If you outsource shipping, that is one more supplier whose security you should be asking about, not assuming.

Source: BleepingComputer

🤖 AI

The open-weights fight spills into the open, and Anthropic sits it out

The policy debate over whether Washington should restrict downloadable "open-weight" AI models turned into a public loyalty test this weekend. On Friday, Nvidia CEO Jensen Huang circulated a letter titled "Open Weights and American AI Leadership," arguing that the government should not impose premature restrictions on models whose trained parameters are published openly, on the grounds that open weights speed innovation, strengthen cybersecurity, and support American technological leadership. The initial signatories included Nvidia, Microsoft, Meta, IBM, Dell, Mistral, Hugging Face, Replit, and Perplexity, among others.

The list then doubled to 50 signatories within a day, adding Google, AMD, Cisco, Cloudflare, GitHub, and others, with OpenAI joining after Sam Altman posted that he wanted the US to "win with both open-weight and proprietary models." The conspicuous holdouts were Amazon and Anthropic. For Anthropic, which has built its brand around AI safety and has been the most vocal major lab about the risks of powerful models, staying off a letter opposing model restrictions is a coherent, if lonely, position. The backdrop is a live debate in Washington over how to handle powerful Chinese open-weight models.

In short: Nvidia's "Open Weights and American AI Leadership" letter opposing restrictions on downloadable models doubled to 50 signatories in a day, with OpenAI joining but Anthropic and Amazon staying out.

What it means for your business: How Washington comes down on open-weight models affects which AI tools stay legally available and how much of your stack you can self-host versus rent from an API. If you rely on open models for cost or data-control reasons, this is a policy fight worth tracking.

My take: This is really a proxy war over two visions of AI: open and widely distributed versus closed and tightly controlled. Reasonable people land in different places, and Anthropic's absence is consistent with its long-standing safety posture rather than a surprise. For business owners the takeaway is simpler: do not assume today's open-model availability is permanent, and keep an eye on where the rules settle.

Source: CNBC

That is your AI and IT news for the weekend of July 24 to 26, 2026. If you run Java applications or PTC Windchill, the two unpatched-and-exploited stories at the top of the security section are the ones to act on first.