AI and IT News Recap: July 20, 2026: A Routing Bug Cracks WordPress Wide Open, Anthropic Courts Meta for 0 Billion, and Fable 5's Long GoodbyeBy Noah Smith, Owner & Consultant, KeyChange Technologies · July 20, 2026
Over the weekend the news was less about a shiny new model and more about the plumbing underneath your business getting stress-tested. This edition of the AI and IT news covers a WordPress flaw that turns any anonymous visitor into an administrator, two fresh attacks aimed squarely at developers, Anthropic shopping for compute in an unexpected place, and Fable 5's drawn-out goodbye finally arriving.
📌 The AI and IT news at a glance
- WordPress Core cracked wide open: A pre-auth "wp2shell" chain hands unauthenticated attackers full remote code execution, and it is already being exploited.
- Poisoned RubyGems target your laptop, not your CI: The "SleeperGem" campaign fakes Microsoft's Git Credential Manager and deliberately skips build servers.
- Seven booby-trapped npm packages hit Vite developers: "ViteVenom" hides a remote access trojan behind blockchain-based command-and-control.
- Anthropic courts Meta for up to 0 billion in compute: The two rivals are in early talks over a two-year cloud lease.
- Fable 5's long goodbye lands today: Included access ended July 19; metered credits and a new tier split begin July 20.
- OpenAI folds Codex into a unified ChatGPT desktop app: One switcher now covers Chat, Work, and Codex across your devices.
🔝 Top story: A single routing mix-up cracks WordPress Core wide open
A newly disclosed exploit chain nicknamed "wp2shell" lets an unauthenticated attacker run code on essentially any WordPress site, with no plugins or theme required. It combines two flaws in WordPress core itself: CVE-2026-63030, a REST API batch-route confusion bug, and CVE-2026-60137, a SQL injection. On their own each is limited, but chained they are brutal. An anonymous request to the REST batch endpoint (something like /wp-json/batch/v1) is crafted so WordPress dispatches one sub-request using the handler meant for a different one. That misrouting sidesteps the authentication checks that would normally block access, reaching a vulnerable database query path, and from there an attacker moves from SQL injection to full database compromise to remote code execution.
The reason this one deserves the top slot is timing. The full technical writeup and a working proof-of-concept went public on Friday, and researchers reported early signs of in-the-wild exploitation within about 24 hours. WordPress shipped fixes in versions 6.8.6, 6.9.5, and 7.0.2 and took the unusual step of force-pushing the update through its auto-update system to all supported installations. That is a strong signal of how seriously the maintainers are taking it.
In short: Two chained WordPress core bugs let anonymous attackers reach remote code execution, a public exploit is circulating, and emergency patches are being force-pushed.
What it means for your business: If you run any public WordPress site, treat this as a same-day patch item and confirm you are on a fixed version, because auto-update can be disabled and attackers are already scanning.
My take: The scary part here is not a clever memory exploit, it is a logic bug in how requests get routed. Those are easy to miss in review and easy to weaponize once someone spots them. Verify the patch actually applied rather than assuming auto-update did its job, and if the site was internet-facing over the weekend, check it for signs of compromise instead of just patching and moving on.
Source: The Hacker News
🛡️ IT and security
Poisoned RubyGems that want your laptop, not your build server
A supply-chain campaign researchers named "SleeperGem" pushed malicious versions of three RubyGems packages to RubyGems.org over July 18 and 19. The headliner is a gem called git_credential_manager that impersonates Microsoft's official Git Credential Manager, joined by tampered versions of Dendreo and a fastlane testing plugin. What makes this one notable is its restraint. The malware pulls a second stage from an attacker-controlled Forgejo host, then checks for roughly thirty environment variables that continuous integration platforms set, including GITHUB_ACTIONS, GITLAB_CI, and CIRCLECI. If it detects any of them, it quietly exits and does nothing.
The logic is deliberate. Ephemeral CI runners hold little of value and carry a high chance of detection, so the attackers want to land on a real developer's machine, where they drop a native background daemon and install persistence. Defenders spotted it because gems that had not shipped a release in years suddenly published new versions minutes apart, and one version appeared with no matching tag in its source repository.
In short: Three malicious RubyGems, one posing as Microsoft's Git Credential Manager, deliberately skip CI systems to backdoor developer workstations.
What it means for your business: Anyone on your team who installs Ruby dependencies is a target, so pin versions, review lockfiles, and be suspicious of long-dormant packages that suddenly update.
My take: The CI-evasion trick is the tell that these attackers are getting more thoughtful, not less. They are hunting for credentials and long-lived tokens that live on human machines. A developer laptop with cloud keys is a far richer prize than a throwaway build container, and most endpoint tooling watches servers more closely than laptops.
Source: The Hacker News
Seven booby-trapped npm packages hit the Vite ecosystem
Checkmarx disclosed a separate developer-focused campaign it calls "ViteVenom," made up of seven malicious npm packages aimed at users of the popular Vite frontend tooling. The packages used scoped names closely resembling the legitimate @vitejs namespace, banking on a distracted developer grabbing the wrong one. Once installed, they reach out through a four-tier, blockchain-based command-and-control setup to deliver a remote access trojan capable of reverse shell access, credential harvesting, file exfiltration, and planting a persistent backdoor.
Researchers tie the activity to a threat actor they track as "SuccessKey," with linked cryptocurrency wallets active as far back as late February. The seven packages were uploaded between late June and early July and pulled hundreds of downloads before being caught, and Checkmarx frames the effort as an evolution of an earlier operation called ChainVeil.
In short: Seven npm packages mimicking Vite's official namespace used blockchain-based command-and-control to drop a remote access trojan on developer machines.
What it means for your business: Typo-squatting on trusted namespaces keeps working, so double-check package names and scopes before installing and consider a tool that flags known-malicious packages at install time.
My take: Using a blockchain for command-and-control is not attackers being trendy, it is them making takedowns harder because there is no single server to seize. Between this and the RubyGems campaign in the same weekend, the pattern is clear: the package registries your developers trust are now a primary front door, and "it came from the official registry" is no longer reassurance.
Source: The Hacker News
🤖 AI
Anthropic is in talks to lease up to 0 billion in compute from Meta
In one of the odder pairings of the year, Anthropic is in early discussions to lease computing power from Meta in an arrangement reportedly worth up to
0 billion over two years. Anthropic is said to have floated the idea in June, with payments spread in monthly increments and either side free to walk away. For Meta, renting out capacity would mark the start of the cloud-services business Mark Zuckerberg has been hinting at, notable given Meta is guiding toward as much as 45 billion in capital spending this year.
For Anthropic, it is the latest in a string of large compute commitments as it scrambles to keep up with demand for paid access. The talks come after a reported deal to use SpaceX capacity worth roughly $45 billion over three years. Nothing is signed, and both companies declined to comment, so treat this as direction rather than done.
In short: Anthropic and Meta are in early talks over a two-year compute lease that could be worth up to
0 billion.
What it means for your business: The scale of these compute deals is why AI pricing stays volatile, and it is a reminder that the vendor you depend on may be leaning on a competitor's hardware behind the scenes.
My take: Two rivals striking a supplier relationship tells you compute scarcity now outranks competitive awkwardness. For anyone building on these models, it is worth remembering that your provider's costs, and therefore your prices and rate limits, ultimately trace back to who can secure the most silicon.
Source: CNBC
Fable 5's long goodbye finally arrives
After a run of last-minute extensions, Anthropic's included access to Claude Fable 5 reached its end. The final reprieve, which came with a 50 percent rate-limit boost, expired on July 19 at 11:59 pm Pacific. Starting today, July 20, Fable 5 usage moves to prepaid usage credits priced at
0 per million input tokens and $50 per million output tokens. The plan structure also splits: Max and Team Premium subscribers get Fable 5 back at 50 percent of their weekly limits, while Pro and Team Standard move to credit-based access with a one-time 00 credit to soften the transition.
If your team has been leaning on Fable 5 through a subscription, this is the change that has been threatened for weeks actually taking effect, so it is worth checking which tier you are on and what it now costs.
In short: Included Fable 5 access ended July 19, and metered credits at
0/$50 per million tokens plus a new tier split begin July 20.
What it means for your business: Heavy Fable 5 users on Pro or Team Standard should model their monthly token spend now, because usage that felt free under a subscription can add up quickly on metered pricing.
My take: The repeated extensions were never really about generosity, they were about easing people onto a pricing model that reflects what top-tier inference actually costs. The useful move is not to complain about the change but to measure your real usage, because most teams have no idea how many tokens they burn until a meter is attached.
Source: ExplainX
🧰 New tools for builders
OpenAI folds Codex into a unified ChatGPT desktop app
OpenAI pushed a redesigned desktop app to all plans on macOS and Windows. The headline change is a single global switcher that lets you jump between ChatGPT and Codex, and within ChatGPT, choose between Chat for quick questions and Work for end-to-end tasks. Recents now unify chat and work threads in one list you can sort, filter, and pin, your existing Projects show up inside the app, and Work conversations sync through the cloud so you can start on one device and pick up on another.
It is an incremental, quality-of-life update rather than a new capability, but the direction matters. OpenAI is clearly trying to make the desktop the default surface where coding and general assistance live side by side rather than in separate tools.
In short: OpenAI's redesigned desktop app adds a Chat, Work, and Codex switcher, unified Recents, in-app Projects, and cross-device sync, now live on macOS and Windows.
What it means for your business: If your team already pays for ChatGPT, the built-in Codex access and cross-device continuity may reduce the need for separate coding tools, so it is worth a look before renewing another subscription.
My take: Consolidating chat, agentic work, and coding into one app is a quiet bid to become the place you spend your whole day. That is convenient, and it also deepens lock-in, so enjoy the smoother workflow while keeping an honest eye on how much of your stack sits with a single vendor.
Source: Neowin
That is the AI and IT news for July 20, 2026. For last week's rundown, see the July 17 recap. Patch your WordPress, watch what your developers install, and keep an eye on your Fable 5 bill.
By Noah Smith, Owner & Consultant, KeyChange Technologies · July 20, 2026
Over the weekend the news was less about a shiny new model and more about the plumbing underneath your business getting stress-tested. This edition of the AI and IT news covers a WordPress flaw that turns any anonymous visitor into an administrator, two fresh attacks aimed squarely at developers, Anthropic shopping for compute in an unexpected place, and Fable 5's drawn-out goodbye finally arriving.
📌 The AI and IT news at a glance
- WordPress Core cracked wide open: A pre-auth "wp2shell" chain hands unauthenticated attackers full remote code execution, and it is already being exploited.
- Poisoned RubyGems target your laptop, not your CI: The "SleeperGem" campaign fakes Microsoft's Git Credential Manager and deliberately skips build servers.
- Seven booby-trapped npm packages hit Vite developers: "ViteVenom" hides a remote access trojan behind blockchain-based command-and-control.
- Anthropic courts Meta for up to 0 billion in compute: The two rivals are in early talks over a two-year cloud lease.
- Fable 5's long goodbye lands today: Included access ended July 19; metered credits and a new tier split begin July 20.
- OpenAI folds Codex into a unified ChatGPT desktop app: One switcher now covers Chat, Work, and Codex across your devices.
🔝 Top story: A single routing mix-up cracks WordPress Core wide open
A newly disclosed exploit chain nicknamed "wp2shell" lets an unauthenticated attacker run code on essentially any WordPress site, with no plugins or theme required. It combines two flaws in WordPress core itself: CVE-2026-63030, a REST API batch-route confusion bug, and CVE-2026-60137, a SQL injection. On their own each is limited, but chained they are brutal. An anonymous request to the REST batch endpoint (something like /wp-json/batch/v1) is crafted so WordPress dispatches one sub-request using the handler meant for a different one. That misrouting sidesteps the authentication checks that would normally block access, reaching a vulnerable database query path, and from there an attacker moves from SQL injection to full database compromise to remote code execution.
The reason this one deserves the top slot is timing. The full technical writeup and a working proof-of-concept went public on Friday, and researchers reported early signs of in-the-wild exploitation within about 24 hours. WordPress shipped fixes in versions 6.8.6, 6.9.5, and 7.0.2 and took the unusual step of force-pushing the update through its auto-update system to all supported installations. That is a strong signal of how seriously the maintainers are taking it.
In short: Two chained WordPress core bugs let anonymous attackers reach remote code execution, a public exploit is circulating, and emergency patches are being force-pushed.
What it means for your business: If you run any public WordPress site, treat this as a same-day patch item and confirm you are on a fixed version, because auto-update can be disabled and attackers are already scanning.
My take: The scary part here is not a clever memory exploit, it is a logic bug in how requests get routed. Those are easy to miss in review and easy to weaponize once someone spots them. Verify the patch actually applied rather than assuming auto-update did its job, and if the site was internet-facing over the weekend, check it for signs of compromise instead of just patching and moving on.
Source: The Hacker News
🛡️ IT and security
Poisoned RubyGems that want your laptop, not your build server
A supply-chain campaign researchers named "SleeperGem" pushed malicious versions of three RubyGems packages to RubyGems.org over July 18 and 19. The headliner is a gem called git_credential_manager that impersonates Microsoft's official Git Credential Manager, joined by tampered versions of Dendreo and a fastlane testing plugin. What makes this one notable is its restraint. The malware pulls a second stage from an attacker-controlled Forgejo host, then checks for roughly thirty environment variables that continuous integration platforms set, including GITHUB_ACTIONS, GITLAB_CI, and CIRCLECI. If it detects any of them, it quietly exits and does nothing.
The logic is deliberate. Ephemeral CI runners hold little of value and carry a high chance of detection, so the attackers want to land on a real developer's machine, where they drop a native background daemon and install persistence. Defenders spotted it because gems that had not shipped a release in years suddenly published new versions minutes apart, and one version appeared with no matching tag in its source repository.
In short: Three malicious RubyGems, one posing as Microsoft's Git Credential Manager, deliberately skip CI systems to backdoor developer workstations.
What it means for your business: Anyone on your team who installs Ruby dependencies is a target, so pin versions, review lockfiles, and be suspicious of long-dormant packages that suddenly update.
My take: The CI-evasion trick is the tell that these attackers are getting more thoughtful, not less. They are hunting for credentials and long-lived tokens that live on human machines. A developer laptop with cloud keys is a far richer prize than a throwaway build container, and most endpoint tooling watches servers more closely than laptops.
Source: The Hacker News
Seven booby-trapped npm packages hit the Vite ecosystem
Checkmarx disclosed a separate developer-focused campaign it calls "ViteVenom," made up of seven malicious npm packages aimed at users of the popular Vite frontend tooling. The packages used scoped names closely resembling the legitimate @vitejs namespace, banking on a distracted developer grabbing the wrong one. Once installed, they reach out through a four-tier, blockchain-based command-and-control setup to deliver a remote access trojan capable of reverse shell access, credential harvesting, file exfiltration, and planting a persistent backdoor.
Researchers tie the activity to a threat actor they track as "SuccessKey," with linked cryptocurrency wallets active as far back as late February. The seven packages were uploaded between late June and early July and pulled hundreds of downloads before being caught, and Checkmarx frames the effort as an evolution of an earlier operation called ChainVeil.
In short: Seven npm packages mimicking Vite's official namespace used blockchain-based command-and-control to drop a remote access trojan on developer machines.
What it means for your business: Typo-squatting on trusted namespaces keeps working, so double-check package names and scopes before installing and consider a tool that flags known-malicious packages at install time.
My take: Using a blockchain for command-and-control is not attackers being trendy, it is them making takedowns harder because there is no single server to seize. Between this and the RubyGems campaign in the same weekend, the pattern is clear: the package registries your developers trust are now a primary front door, and "it came from the official registry" is no longer reassurance.
Source: The Hacker News
🤖 AI
Anthropic is in talks to lease up to
0 billion in compute from MetaIn one of the odder pairings of the year, Anthropic is in early discussions to lease computing power from Meta in an arrangement reportedly worth up to
0 billion over two years. Anthropic is said to have floated the idea in June, with payments spread in monthly increments and either side free to walk away. For Meta, renting out capacity would mark the start of the cloud-services business Mark Zuckerberg has been hinting at, notable given Meta is guiding toward as much as45 billion in capital spending this year.For Anthropic, it is the latest in a string of large compute commitments as it scrambles to keep up with demand for paid access. The talks come after a reported deal to use SpaceX capacity worth roughly $45 billion over three years. Nothing is signed, and both companies declined to comment, so treat this as direction rather than done.
In short: Anthropic and Meta are in early talks over a two-year compute lease that could be worth up to
0 billion.What it means for your business: The scale of these compute deals is why AI pricing stays volatile, and it is a reminder that the vendor you depend on may be leaning on a competitor's hardware behind the scenes.
My take: Two rivals striking a supplier relationship tells you compute scarcity now outranks competitive awkwardness. For anyone building on these models, it is worth remembering that your provider's costs, and therefore your prices and rate limits, ultimately trace back to who can secure the most silicon.
Source: CNBC
Fable 5's long goodbye finally arrives
After a run of last-minute extensions, Anthropic's included access to Claude Fable 5 reached its end. The final reprieve, which came with a 50 percent rate-limit boost, expired on July 19 at 11:59 pm Pacific. Starting today, July 20, Fable 5 usage moves to prepaid usage credits priced at
0 per million input tokens and $50 per million output tokens. The plan structure also splits: Max and Team Premium subscribers get Fable 5 back at 50 percent of their weekly limits, while Pro and Team Standard move to credit-based access with a one-time00 credit to soften the transition.If your team has been leaning on Fable 5 through a subscription, this is the change that has been threatened for weeks actually taking effect, so it is worth checking which tier you are on and what it now costs.
In short: Included Fable 5 access ended July 19, and metered credits at
0/$50 per million tokens plus a new tier split begin July 20.What it means for your business: Heavy Fable 5 users on Pro or Team Standard should model their monthly token spend now, because usage that felt free under a subscription can add up quickly on metered pricing.
My take: The repeated extensions were never really about generosity, they were about easing people onto a pricing model that reflects what top-tier inference actually costs. The useful move is not to complain about the change but to measure your real usage, because most teams have no idea how many tokens they burn until a meter is attached.
Source: ExplainX
🧰 New tools for builders
OpenAI folds Codex into a unified ChatGPT desktop app
OpenAI pushed a redesigned desktop app to all plans on macOS and Windows. The headline change is a single global switcher that lets you jump between ChatGPT and Codex, and within ChatGPT, choose between Chat for quick questions and Work for end-to-end tasks. Recents now unify chat and work threads in one list you can sort, filter, and pin, your existing Projects show up inside the app, and Work conversations sync through the cloud so you can start on one device and pick up on another.
It is an incremental, quality-of-life update rather than a new capability, but the direction matters. OpenAI is clearly trying to make the desktop the default surface where coding and general assistance live side by side rather than in separate tools.
In short: OpenAI's redesigned desktop app adds a Chat, Work, and Codex switcher, unified Recents, in-app Projects, and cross-device sync, now live on macOS and Windows.
What it means for your business: If your team already pays for ChatGPT, the built-in Codex access and cross-device continuity may reduce the need for separate coding tools, so it is worth a look before renewing another subscription.
My take: Consolidating chat, agentic work, and coding into one app is a quiet bid to become the place you spend your whole day. That is convenient, and it also deepens lock-in, so enjoy the smoother workflow while keeping an honest eye on how much of your stack sits with a single vendor.
Source: Neowin
That is the AI and IT news for July 20, 2026. For last week's rundown, see the July 17 recap. Patch your WordPress, watch what your developers install, and keep an eye on your Fable 5 bill.