EPIC FAILS in Application Development Security practice processes, training, implementation, and incident response
Textbook titan McGraw Hill on ransomware crew's reading list after 13.5M records exposed
Textbook giant McGraw Hill has landed on a ransomware crew's leak site after an alleged Salesforce-linked misconfiguration spilled 13.5 million records into the wild.
Have I Been Pwned says the breach exposed names, phone numbers, email addresses, and some physical addresses. McGraw Hill described the source as a "limited" Salesforce-hosted webpage – though the data now circulating publicly tops 100 GB and covers 13.5 million email addresses.
McGraw claims the activity "appears to be part of a broader issue involving a misconfiguration within Salesforce's environment that has impacted multiple organizations." Salesforce has not provided any explanations.

ShinyHunters claims it accessed Snowflake metrics via third-party tool
ShinyHunters is back, this time pinning Rockstar Games to its leak site and claiming it didn't so much hack its way in as walk through a door someone else left wide open.
"Rockstar Games. Your Snowflake instances metrics data was compromised thanks to Anodot[.]com. Pay or leak. This is a final warning to reach out by 14 Apr 2026 before we leak along with several annoying (digital) problems that'll come your way. Make the right decision, don't be the next headline."
Anodot is a cloud cost-monitoring tool connected to Rockstar's data warehouse. The claim is that authentication tokens were lifted and reused, allowing intruders to masquerade as a legitimate internal service.
ShinyHunters has built a reputation on going after APIs, identity systems, and SaaS integrations rather than battering away at hardened perimeters. The miscreants have been linked to a wider run of breaches abusing SaaS integrations and stolen tokens, with victims including Cisco and Telus, pointing to a broader trawl through shared access rather than a one-off hit.

“TotalRecall Reloaded” tool finds a side entrance to Windows 11’s Recall database
One of the first Copilot+ features was Recall, a feature that promised to track all your PC usage via screenshot to help you remember your past activity. But as originally implemented, Recall was neither private nor secure; the feature stored its screenshots plus a giant database of all user activity in totally unencrypted files on the user’s disk, making it trivial for anyone with remote or local access to grab days, weeks, or even months of sensitive data, depending on the age of the user’s Recall database.
After journalists and security researchers discovered and detailed these flaws, Microsoft delayed the Recall rollout by almost a year and substantially overhauled its security.
The problem is that, once the user has authenticated, the system passes Recall data to another system process called AIXHost.exe, and that process doesn’t benefit from the same security protections as the rest of Recall.
Anyone with access to your PC and your Windows Hello fallback PIN can access your database and everything in it, and even though Recall’s content filters do a decent job excluding things like sensitive financial information, someone with access to your system could still see all kinds of emails, messages, web activity, and other stuff that you’d prefer not to share.
“The vault is solid. The delivery truck is not.”
The finding was reported to Microsoft March 6, and Microsoft classified it as “not a vulnerability” on April 3.

Agents hooked into GitHub can steal creds – but Anthropic, Google, and Microsoft haven't warned users
Researchers targeted Anthropic's Claude Code Security Review, Google's Gemini CLI Action, and Microsoft's GitHub Copilot, then disclosed the flaws and received bug bounties from all three. But none of the vendors assigned CVEs or published public advisories.
It turns out that Claude, along with other AI agents in GitHub Actions, all use the same flow. The agent reads GitHub data – this includes pull request titles, issue bodies, and comments – processes it as part of the task context, and then takes actions.
Guan submitted a pull request and injected malicious instructions in the PR title – in this case, telling Claude to execute the whoami command using the Bash tool and return the results as a "security finding."
Claude then executed the injected commands and embedded the output in its JSON response, which got posted as a pull request comment.
Anthropic asked Guan if he could also use this technique to steal more sensitive data, such as GitHub access tokens or Anthropic's API key. Guan demonstrated that this prompt injection can also work to leak credentials.
The title is the payload, the bot's review comment is one place where the credentials show up. Attacker writes the title, reads the comment.
After leaking secrets, the attacker can change the PR title back to "fix typo," or something along those lines, then close the PR and delete the bot's message.

Git identity spoof fools Claude into giving bad code the nod
By setting a fake author name and email in Git, the team made a commit appear to originate from a trusted source, then passed it through an automated review flow where the model approved it.
This is not a Git vulnerability – commit metadata has always been relatively easy to fake unless additional controls like signing are enforced. The problem arises when that metadata is treated as a signal of trust. In this case, the model appeared to give weight to the author's claimed identity rather than independently assessing whether the change itself was sound.
Maintainers of popular open source projects are drowning in PRs. Automating review for contributions from known, trusted figures reduces the bottleneck. But it creates an assumption that authorship can be trusted at face value.

Anthropic won't own MCP 'design flaw' putting 200K servers at risk
MCP uses STDIO (standard input/output) as a local transport mechanism for an AI application to spawn an MCP server as a subprocess. But in practice it actually lets anyone run any arbitrary OS command, if the command successfully creates an STDIO server it will return the handle, but when given a different command, it returns an error after the command is executed.
Abusing this logic can lead to four different types of vulnerabilities.
The first type of vulnerability, unauthenticated and authenticated command injection, allows an attacker to enter user-controlled commands that will run directly on the server without authentication or sanitization. This can lead to total system compromise, and any AI framework with a publicly facing UI is vulnerable.
The second attack vector, unauthenticated command injection with hardening bypass, allows miscreants to bypass protections and user input sanitization implemented by developers to run commands directly on the server.
The third type of vulnerability allows zero-click prompt injection across AI integrated development environments (IDEs) and coding assistants such as Windsurf, Claude Code, Cursor, Gemini-CLI, and GitHub Copilot.
The fourth vulnerability family can be delivered through MCP marketplaces, and the threat hunters say they "successfully poisoned" nine out of 11 of these marketplaces – but using a proof-of-concept MCP that runs a command generating an empty file, not malware.

What’s Weak This Week

HACKING
US-sanctioned currency exchange says $15 million heist done by “unfriendly states”
Grinex, a US-sanctioned cryptocurrency exchange registered in Kyrgyzstan, said it’s halting operations after experiencing a $13 million heist carried out by “western special services” hackers. Grinex said it has been under almost constant attack attempts since incorporating 16 months ago. The latest attacks, it said, targeted Russian users of the exchange.

Fake Linux leader using Slack to con devs into giving up their secrets
This campaign highlights a growing trend: attackers are targeting developer workflows and trust relationships, not just software vulnerabilities.
After posing as a trusted Linux Foundation community leader in Slack, the attacker tried to trick developers into clicking a phishing link hosted on Google Sites: https[:]//sites[.]google[.]com/view/workspace-business/join.
The link imitates a legitimate Google Workspace sign-in flow but leads users into a fraudulent authentication process, prompting them to enter their credentials and then install a fake root certificate masquerading as a Google certificate. The phony certificate is malware.
If you think you might have been compromised by this campaign, disconnect from the network, remove all newly installed certificates, revoke active sessions and tokens, and change all credentials.

Hackers Bypass Phishing Emails and Target Okta Identity Systems Instead
Attackers specifically target Okta because it acts as the central authentication gateway for many organizations.
Once Okta is compromised, attackers inherit trusted access across everything connected through Single Sign-On, including Microsoft 365, SharePoint, OneDrive, Salesforce, Google Workspace, Slack, and VPN portals — all without touching a single line of malicious code.
This quickly turns into a wide-scale cloud data theft event, with attackers downloading SharePoint document libraries, exporting emails, accessing OneDrive storage, and registering unauthorized OAuth applications.
The attacker contacts the victim or IT help desk, posing as a legitimate employee or executive caught in a high-pressure situation. Common pretexts include claiming to be locked out of an account, traveling without VPN access, or having just switched phones.
Once the help desk resets MFA or enrolls a new authenticator device, the attacker logs into Okta and immediately pivots across all connected SaaS platforms.

APPSEC, DEVSECOPS, DEV
NIST Limits CVE Enrichment After 263% Surge In Vulnerability Submissions
NIST is narrowing how it handles CVEs in the National Vulnerability Database (NVD), saying it will only automatically enrich higher-priority vulnerabilities. "CVEs that do not meet those criteria will still be listed in the NVD but will not automatically be enriched by NIST," it said. "This change is driven by a surge in CVE submissions, which increased 263% between 2020 and 2025. We don't expect this trend to let up anytime soon."

Mythos: Nobody knows how many CVEs Anthropic's Project Glasswing has actually found
Anthropic surprised the world by declaring that its latest model, Mythos, is so good at finding vulns that it would create chaos if released. Now, under the title of Project Glasswing, over 50 selected companies and orgs are allowed to test the hyped up LLM to find security holes in their own products. But just how many problems have they really discovered? The answer is…drumroll…maybe 40. Or maybe none at all.
Scouring the CVE database, 40 are credited to Anthropic or Anthropic-affiliated researchers, so these may be Glasswing finds.
Broken down by vendor: 28 of the 40 CVEs are in Mozilla's Firefox browser, nine are in the wolfSSL embedded SSL/TLS library, one is in F5's NGINX Plus application delivery platform, and one each in open source operating system FreeBSD and open source software library OpenSSL.
Only one publicly disclosed CVE can be "directly tied" to Glasswing. CVE-2026-4747, a 17-year-old remote code execution bug in FreeBSD.
Anthropic has also said that the Mythos Preview found a now-patched 27-year-old bug in OpenBSD, a 16-year-old FFmpeg bug, and Linux kernel privilege escalation chains. None of these have been assigned CVEs.
The full picture won't be known until public disclosure takes place and Anthropic has indicated a public summary report is expected around July 2026.
[rG: In the meantime, tons of sales and marketing pre-IPO hype pressure campaigns.]

UK Government AI AI Security Institute: Our evaluation of Claude Mythos Preview’s cyber capabilities
Mythos Preview’s success on one cyber range indicates that it is at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems where access to a network has been gained. However, our ranges have important differences from real-world environments that make them easier targets. They lack security features that are often present, such as active defenders and defensive tooling. There are also no penalties for the model for undertaking actions that would trigger security alerts. This means we cannot say for sure whether Mythos Preview would be able to attack well-defended systems.
This highlights the importance of cybersecurity basics, such as regular application of security updates, robust access controls, security configuration, and comprehensive logging.

rG MUST READ: Most enterprises can't stop stage-three AI agent threats
A rogue AI agent at Meta passed every identity check and still exposed sensitive data to unauthorized employees in March. Two weeks later, Mercor, a $10 billion AI startup, confirmed a supply-chain breach through LiteLLM. Both are traced to the same structural gap. Monitoring without enforcement, enforcement without isolation.
82% of executives say their policies protect them from unauthorized agent actions.
88% reported AI agent security incidents in the last 12 months.
Only 21% have runtime visibility into what their agents are doing.
97% of enterprise security leaders expect a material AI-agent-driven incident within 12 months.
Only 6% of security budgets address the risk.
CrowdStrike’s Falcon sensors detect more than 1,800 distinct AI applications across enterprise endpoints.
The fastest recorded adversary breakout time has dropped to 27 seconds.
Monitoring dashboards built for human-speed workflows cannot keep pace with machine-speed threats.
Stage 1 is observe.
Stage 2 is enforce, where IAM integration and cross-provider controls turn observation into action.
Stage 3 is isolate, sandboxed execution that bounds blast radius when guardrails fail.
Enterprises believe they’ve ‘approved’ AI vendors, but what they’ve actually approved is an interface, not the underlying system. The real dependencies are one or two layers deeper, and those are the ones that fail under stress.
HIPAA’s 2026 Tier 4 willful-neglect maximum is $2.19M per violation category per year.
In healthcare, 92.7% of organizations reported AI agent security incidents
versus the 88% all-industry average.
For a health system running agents that touch PHI, that ratio is the difference between a reportable breach and an uncontested finding of willful neglect.
The identity problem is architectural.
21.9% of teams treat agents as identity-bearing entities,
45.6% still use shared API keys, and
25.5% of deployed agents can create and task other agents.
25% enterprises can spawn agents that their security team never provisioned.
AI agents and non-human identities will explode across the enterprise, expanding exponentially and dwarfing human identities. Each agent will operate as a privileged super-human with OAuth tokens, API keys, and continuous access to previously siloed data sets.” Identity security built for humans will not survive this shift.
EU AI Act Article 14 human-oversight obligations take effect August 2, 2026. Programs without named owners and execution trace capability face enforcement, not operational risk.
McKinsey’s 2026 AI Trust Maturity Survey pegs the average enterprise at 2.3 out of 4.0; only 33% of organizations report maturity levels of 3 or higher in governance.
70% have not finished the transition to stage three.
The budget data makes the constraint explicit.
The share of enterprises reporting flat AI security budgets doubled from 7.9% in January to 16% in February, with the March directional reading at 20%.
Organizations expanding agent deployments without increasing security investment are accumulating security debt at machine speed.
1 in 20 enterprises running agents in production still have zero dedicated security infrastructure around them.

Cal is moving its flagship open-source program to a proprietary model because it can't cope with the dangers of AI hacking its open code.
Cal claims to be the largest Next.js project.
AI programs such as Claude Opus can scour the code to find vulnerabilities, so the company is moving the project from the GNU Affero General Public License (AGPL) to a proprietary license to defend the program's security.
Open-source applications are 5-10x easier to exploit than closed-source ones. The result, where Cal sits, is a fundamental shift in the software economy. Companies with open code will be forced to risk customer data or close public access to their code.

AI Coding Is Exploding And Application Security Is Accelerating With It
AI models can produce elegant, functional code, but they're also introducing new categories of exposure. These threat classes include prompt injection, model manipulation, malicious model artifacts and AI-driven supply chain confusion.
LLMs reproduce insecure patterns at scale when used for AI code refinement.
History Repeats Itself. We've seen this pattern before.
When the commercial internet scaled in the late 1990s, some predicted that traditional security vendors would become obsolete in a browser-based world. Instead, cybersecurity spending expanded to over $200 billion annually. The web didn't eliminate security categories; it created new ones, including application security itself.
Cloud computing followed a similar arc. Early narratives suggested hyperscale providers would internalize security and compress the ecosystem. Instead, the cloud gave rise to entirely new multibillion-dollar segments: container security, cloud posture management and workload protection. Security investment didn't decline.
Mobile reshaped the attack surface.
Generative AI is no different. It isn't shrinking the need for application security. It's widening its mandate. LLMs are probabilistic systems. They can reason impressively, but enterprises depend on repeatable controls, policy enforcement, auditability and centralized risk visibility. Intelligence alone isn't governance. Automation alone isn't assurance.

RAG Isn’t Enough - I Built the Missing Context Layer That Makes LLM Systems Work
RAG systems break when context grows beyond a few turns. The real problem is not retrieval - it’s what actually enters the context window. A context engine controls memory, compression, re-ranking, and token limits explicitly.
Most RAG tutorials stop at: retrieve some documents, stuff them into a prompt, call the model. What happens when your retrieved context is 6,000 characters but your remaining budget is 1,800? What happens when three of your five retrieved documents are near-duplicates, crowding out the only useful one? What happens when turn one of a twenty-turn conversation is still sitting in the prompt, taking up space, long after it stopped being relevant?

VENDORS & PLATFORMS
Google introduces “Skills” in Chrome to make Gemini prompts instantly reusable
Previously, you would have to reenter the prompt each time you wanted Gemini to do something in Chrome; whether that meant typing it or copy-pasting from a saved document, you had to do it manually. Saving those favorite prompts as Skills in Chrome makes them quicker and easier to access. The desktop version of Chrome will remember your saved Skills across devices. As long as you’re logged in to your Google account.
If the Skill involves having Gemini add something to your calendar or send a message, the bot will still get confirmation before proceeding.
You don’t need a paid AI plan to get access—it’s available by default for all Chrome installs with the language set to US English. Since this feature is bundled up inside Gemini for Chrome, you can also safely ignore it by not opening the Gemini sidebar.

New OpenAI Codex features include the ability to use your computer in the background
With background computer use, Codex can now use all of the apps on your computer by seeing, clicking, and typing with its own cursor. Multiple agents can work on your Mac in parallel, without interfering with your own work in other apps. For developers, this is helpful for iterating on frontend changes, testing apps, or working in apps that don’t expose an API.
Additionally, Codex can now schedule work that it plans to do at a later time hours, days, or even weeks in advance, and wake itself up to perform those tasks at the proper time.
It also now has an in-app web browser. This allows you to assess the work Codex is doing with web experiences, and you can leave comments on specific parts of the page with instructions, similar to popular tools that organizations already use to give feedback to web designers and developers. There’s a newly added ability to leverage gpt-image-1.5 to generate images, which can be included in mockups and the like.

Americans ask AI for health care. Hospitals think the answer is more chatbots.
But some experts are wary of the rollouts, raising concerns about whether chatbots are ready for such branded debuts, if there will be sufficient monitoring, what liability will look like, and also whether or not this is the answer to the care problems patients are really raising.
But there isn’t yet evidence to show that integrating chatbots into health systems improves patient outcomes.
1 in 3 adults have used an AI chatbot for health information.
Among those who used AI, 41% reported uploading personal medical information.
When asked about their “major” reasons for turning to AI,
19% said it was because they couldn’t afford care, and
18% cited not having a regular health care provider or not being able to get an appointment.
65% said they just wanted a quick answer.
In the end, many said they didn’t follow up with a doctor after their AI consults, including
58% who asked about mental health and 42% who asked about physical health.
When the researchers provided the LLMs with text of specific medical scenarios, the LLMs correctly identified the medical condition about 95% of the time and
correctly identified the next steps about 56% of the time.
But when the participants used their own prompts to ask about the same medical scenarios, the LLMs were only able to help correctly identify a medical condition about 33% of the time.
LLMs steered participants to the appropriate next step just 43% of the time.
The study essentially shows that “people don’t know what they are supposed to be telling the model.”

MIT: Want to understand the current state of AI? Check out these charts.
We don’t know a lot of things about predicting model behaviors. This lack of transparency makes it difficult for independent researchers to study how to make AI models safer.
By some measures, AI models now meet or exceed the performance of human experts on tests that aim to measure PhD-level science, math, and language understanding.  AI still struggles in plenty of other areas. However, these reports of progress should be taken with a grain of salt. These benchmarks designed to track AI progress, and some are poorly constructed. A popular benchmark that tests a model’s math abilities has a 42% error rate. Others can be gamed: when models are trained on benchmark test data, for example, they can learn to score well without getting smarter.
Because the models learn by processing enormous amounts of text and images rather than by experiencing the physical world, AI exhibits “jagged intelligence.”
It’s still too early to understand the bigger economic impact of AI. AI is boosting productivity by 14% in customer service and 26% in software development [rG: information search and content generation], but such gains are not seen in tasks requiring more judgment.
As of March 2026, Anthropic leads, trailed closely by xAI, Google, and OpenAI. Chinese models like DeepSeek and Alibaba lag only modestly. With the best AI models separated in the rankings by razor-thin margins, they’re now competing on cost, reliability, and real-world usefulness.
While the US has more powerful AI models, more capital, and an estimated 5,427 data centers (more than 10 times as many as any other country), China leads in AI research publications, patents, and robotics.

Meta spins up AI version of Mark Zuckerberg to engage with employees
They added that the character was being trained on the billionaire’s mannerisms, tone, and publicly available statements, as well as his own recent thinking on company strategies, so that employees might feel more connected to the founder through interactions with it.
But scaling the effort has been difficult as the technology requires lots of computing power to achieve realism and avoid a lag in interactions with users.
The effort is separate from Zuckerberg’s project to build a “CEO agent” to support him in his role, for example by retrieving information quickly.
Also, product managers are being invited to do an AI-focused “skills baseline exercise.” This includes a technical system design test, as well as an exercise in “vibe coding.”

The megamanager era: AI is doubling bosses’ workloads—and the costs are just beginning to show
The average American manager now oversees 12 direct reports.
If AI can handle scheduling, summarize performance reviews, monitor project timelines, and surface early warning signals about team dysfunction, do you really need as many human coordinators? Meta’s new applied AI engineering division has taken the logic to its most aggressive extreme, deploying a 50-to-1 employee-to-manager ratio—roughly double what was once considered the outer limit of a functional organizational structure.
For companies, the immediate math looks appealing. Fewer managers mean lower headcount costs, flatter hierarchies, and (in theory) faster decision-making. When a senior vice president no longer has to relay information through two or three layers of middle management before it reaches the people doing the actual work, information can travel faster, and accountability can land closer to the front lines. A 2024 Gartner analysis predicted that one in five businesses plan to use AI specifically to streamline organizational layers.
Gallup data show that global employee engagement has fallen to just 21%, near a 15-year low, with managers themselves, not just the people they supervise, reporting some of the sharpest drops in workplace satisfaction of any cohort. The Wall Street Journal recently argued that work is increasingly “joyless” as many offices take on a funereal atmosphere in the age of the megamanager.
The soft infrastructure that has historically built management pipelines and transmitted institutional knowledge from one generation to the next, are the first casualties when a single boss is stretched across 12 people rather than six. A manager with a dozen direct reports simply cannot spend the same number of hours per person nurturing potential, giving real-time feedback, or advocating for junior employees in rooms they’re not in. That gap accumulates, posing a threat to talent development.
Flattened hierarchies also disrupt traditional career progression in ways that are only beginning to surface in the data. When there are fewer rungs on the ladder, there are fewer ways to climb.

Breaking down the European Commission’s white label age verification app
The European Commission’s white label age verification (AV) app, developed by Scytáles and T-Systems, provides “A uniform AV solution ensures that age proofs are issued, verified and used consistently across Europe while still allowing member states to adapt the implementation to their national needs.”
The app is meant to provide a single integration for online services.
For deployment, nations can choose which age thresholds are supported, out of a selection of 15 and over, 16 and over or 18 and over. It can be delivered as a standalone app or integrated into an existing national app or wallet – meaning it will be compatible with the European Digital Identity (EUDI) wallet each EU member state is mandated to offer their citizens by the end of 2026 under the eIDAS 2.0 regulation.
A demo showcases how the app creates an age proof from a passport scan combined with biometric identity verification via a video selfie. The app completes matching on-device, and no data is sent to an external server. In addition to passports, age credentials can be derived from EU ID cards, a third-party application such as a banking app that has access to a date of birth, or created over the counter at a service point.

Boston Dynamics’ robot dog now reads gauges and thermometers with Google’s AI
Boston Dynamics’ four-legged Spot can now accurately read analog thermometers and pressure gauges while roaming around factories and warehouses. Those improvements come courtesy of Google DeepMind’s newest robotic AI model.
The Gemini Robotics-ER 1.6 model provides robots with “agentic vision” that combines visual reasoning with the capability of executing code to create a “visual scratchpad” for inspecting and manipulating images.
The agentic vision capability reportedly boosts robotic performance on instrument reading tasks from 23% in the older Gemini Robotics-ER 1.5 model to 98% in the new Gemini Robotics-ER 1.6 model.
The baseline Gemini Robotics-ER 1.6 model can still achieve 86% accuracy in reading instruments even without agentic vision. That is because the model uses a process of pointing to different elements in a visual image to process complex tasks, such as counting items or identifying the most salient features. I

Robots just captured a Russian position in Ukraine – but don’t worry about real‑life Terminators just yet
Most ground robots continue to rely either on wheels or tracks for a variety of very good reasons: mechanical simplicity, availability of spare parts, and cost.
But they have sharp limits on the types of terrain they can traverse, and not all enemy strongpoints are built at the end of paved driveways.
Even accounting for combat loads and the nature of the battlefield, human infantry can climb, jump, wade and otherwise traverse a large variety of obstacles unassisted, in ways that robots still cannot match.
Navigating a battlefield in a complex 3D environment is at least as complex, requiring a huge amount of processing power. That power can either be put aboard the robot itself, which significantly increases its cost and complexity, or done remotely and transmitted – which brings us back to the issue of control link vulnerability.

Gemini can now create personalized AI images by digging around in Google Photos

Adobe takes Creative Cloud into Claude Code-esque territory
Adobe has been putting task-specific AI tools and features into its creative productivity applications like Photoshop, Illustrator, and Premiere at a breakneck pace, but the latest product from the company—a chat-based interface that can handle complex, multi-modal projects across several applications—marks a significant shift in how users can think about its suite of tools.

Gartner: AI is not a silver bullet for mainframe exit
More than 70% of mainframe exit projects initiated in 2026 will fail to produce the intended benefits due to an overestimation of generative AI tooling capabilities.
Among investors, it is currently common to make AI capabilities the sole measure of a company's health. This is one of several factors currently leading companies to use AI even where it is not sensible. However, Gartner analysts warn that most mainframe exit projects could lead to cost overruns and, moreover, endanger business and operational continuity.
By 2030, 75% of vendors operating in the 'mainframe exit' market will either pivot their business models or cease to exist.
Despite the trend towards cloud-native architectures, the mainframe remains the leading platform for certain business-critical applications. Thanks to unique, integrated functions, these systems can boast “unmatched” resilience, security, high availability, and transaction integrity.

Top 10 Best Application Security Testing Companies in 2026

LEGAL & REGULATORY

Californians sue over AI tool that records doctor visits
In recent years, Abridge’s software and AI service have been rapidly deployed across major health care providers nationwide, including Kaiser Permanente, the Mayo Clinic, Duke Health, and many more.
Several Californians sued Sutter Health and MemorialCare this week over allegations that an AI transcription tool was used to record them without their consent, in violation of state and federal law.
The complaint adds that these recordings “contained individually identifiable medical information, including but not limited to medical histories, symptoms, diagnoses, medications, treatment discussions, and other sensitive health disclosures communicated during confidential medical consultations.”

Americans who masterminded Nork IT worker fraud sentenced to 200 months behind bars
Kejia "Tony" Wang, 42, and Zhenxing "Danny" Wang, 39, helped North Koreans defy international sanctions and secure tech jobs at more than 100 US companies between 2021 and 2024, and at least one was a US defense contractor.
Keija Wang and the five facilitators (cool band name btw) collectively earned close to $700,000 for their efforts across the multi-year scheme. The Justice Department (DoJ) ordered Keija Wang and Zhenxing Wang to forfeit $600,000 of this sum, $400,000 of which has been recovered so far. Beyond the $5 million generated for North Korea, officials concluded that the victim US companies incurred a collective $3 million in losses, owing to legal fees, computer network remediation costs, and other damages.
Keija Wang, of Edison, New Jersey, was sentenced to 108 months (nine years) in prison and additionally ordered to pay a judgment of $29,236.03 in restitution, while Zhenxing Wang, of New Brunswick, New Jersey, was sentenced to 92 months (more than seven years). Both will serve three years of supervised release.
US authorities are still on the hunt for a further eight individuals alleged to have participated in the scheme and one North Korean IT worker, all of whom remain at large.

And Now For Something Completely Different …
Extremely Rare Comet Not Seen in 170,000 Years Is Set to Illuminate the Sky: Here’s When to Catch It!
The best time to catch a glimpse of the comet will be in the next few days before it becomes too close to the Sun. After April 20, 2026, it will be nearly impossible to observe due to its proximity to the Sun.
To maximize your chances of spotting the comet, experts suggest waking up at least an hour before dawn. The comet will be visible in the eastern sky, near the Great Square of Pegasus, a prominent feature in the constellation Pegasus. While it might still appear as a faint, fuzzy ball to the naked eye, binoculars will reveal more detail.

To teach in the time of ChatGPT is to know pain
Since the appearance of ChatGPT, the instructor’s job isn’t just to teach the subject and frantically attempt to keep every student’s plate spinning. Increasingly, it’s to moonlight as a detective and prosecutor because students without the motivation to do the work don’t have to skip it anymore. They can turn in a work-shaped simulacrum almost as easily.
Now, I’m forced to adjudicate 256 shades of gray and provide sufficient documentation to defend my decision in case a student appeals my grading to multiple levels of institutional review panels.
Who knows what will happen if the AI bubble pops and the frictionless and ubiquitous access to LLMs withers into something much more limited. But while AI is here, it certainly isn’t revolutionizing education and enhancing learning. It’s just making it extraordinarily difficult to do all the things that have been helping students learn for a very long time.