Daily Briefing

May 13, 2026

2026-05-12

19 articles

PwC is deploying Claude to build technology, execute deals, and reinvent enterprise functions for clients

Date not detected

Summary

PwC and Anthropic expand their partnership, with PwC leveraging Claude to build technology for clients, execute deals, and transform enterprise functions.

Key Points

PwC will deploy Claude Code and Cowork from its U.S. team to hundreds of thousands of professionals around the world.
We will establish a joint center of excellence and train and certify 30,000 PwC professionals on Claude.
The collaboration focuses on three areas: building agent technology, AI-powered dealmaking, and reinventing enterprise capabilities.
PwC is launching a new Office of the CFO, based in Claude.
Claude is already operational, reducing delivery times by up to 70% in areas such as insurance underwriting, HR transformation, and cybersecurity.

Notable Quotes & Details

Notable Data / Quotes

$2 trillion (estimated losses from AI transfer systems)
Up to 70% (shortened delivery time)
Insurance underwriting: 10 weeks → 10 days
Security Operations: Hours → Minutes
Dario Amodei, Co-Founder and CEO, Anthropic: “We have been leading the expansion of AI in sectors of the economy where accuracy and reliability are essential, and the results are clear.”
Paul Griggs, PwC US Senior Partner and CEO: “The conversation about AI has shifted from possibility to action. Customers are looking for ways to apply AI in complex business environments that are safe, accountable, and deliver measurable results.”

Intended Audience

Business leaders, executives considering adopting AI technology, IT decision-makers, consultants, and financial services professionals.

Migrating Data Ingestion Systems at Meta Scale

2026-05-12

Summary

Meta successfully migrated its data collection system from a legacy system to a new architecture for expanded data reliability.

Key Points

Meta's data collection system underwent a major migration from a legacy system to a new architecture.
This migration significantly improves efficiency and stability, with a particular focus on social graph data processing.
To ensure a successful migration, stringent validation criteria for data quality, latency, and resource utilization were applied.

Notable Quotes & Details

Notable Data / Quotes

2026/05/12
MySQL
Petabytes of social graph data
100% of the workload

Intended Audience

Data engineers, architects, large-scale system operators

NVIDIA and SAP Bring Trust to Specialized Agents

2026-05-12

Summary

NVIDIA and SAP are collaborating to enhance the security and governance of specialized AI agents within enterprise systems.

Key Points

NVIDIA and SAP have expanded their collaboration to enable enterprises to run specialized AI agents with security and governance controls.
SAP is integrating NVIDIA OpenShell, an open source runtime for developing and deploying secure autonomous AI agents, into SAP Business AI Platform.
SAP engineers co-designed OpenShell with NVIDIA and are contributing to the open source project.
OpenShell provides an isolated execution environment, policy enforcement at the file system and network layers, and infrastructure-level containment to prevent damage when agent logic fails.
This collaboration addresses the needs for trust, vigilance, policy enforcement, and audit trails for autonomous agents in enterprise environments.
NVIDIA OpenShell determines whether the agent task can be safely executed, and the SAP Joule Studio runtime determines whether the task can be performed.

Notable Quotes & Details

Notable Data / Quotes

Announced today at SAP Sapphire
Huang has described AI as a five-layer cake : energy, chips, infrastructure, models and applications.

Intended Audience

Corporate executives, IT experts, developers, and business decision makers interested in AI agent technology adoption and security

JBS Dev: On imperfect data and the AI last mile – from model capability to cost sustainability

2026-05-12

Summary

We'll cover how AI systems handle less-than-perfect data and how the development of AI models will shift toward cost sustainability and portability.

Key Points

AI systems can work with less-than-perfect data, and modern tools address the misconception that they excel at handling low-quality data.
Generative AI and agent AI can be effectively used to process complex and incomplete real-world data.
The future of AI technology will focus less on advancing the capabilities of models and more on achieving cost efficiency and portability to reduce reliance on data centers.

Notable Quotes & Details

Notable Data / Quotes

"It’s a common misconception that your data has to be perfect before you do any of these types of workloads"
"The tooling has never been better than it is now to deal with poor quality data"
"It’s almost remarkable what an LLM can understand on a half-written prompt."
"we started at 20% automated, and then 40%, and then 60, 80%"
"how do we make the cost more sustainable that we don’t have to build data centres at the rate we’re building data centres?"
"The last mile is ‘how do we get these things to run on a laptop or a phone instead of having to run in a data centre?’"

Intended Audience

AI developer, AI project manager, technology executive, AI strategy builder

Laserfiche unveils AI agents for natural language workflows

2026-05-12

Summary

Laserfiche powers content management and workflow automation with the launch of AI agents that perform tasks via natural language prompts.

Key Points

Laserfiche's AI agents use natural language prompts to perform tasks and follow integrated security rules and compliance requirements.
This agent leverages generative LLM inference models to bridge the gap between automated workflows and manual tasks, reducing time and resource consumption.
Accessible through the Smart Chat interface, the agent's capabilities are limited by the user's permissions and restrictions.
Document data analysis allows departments such as legal, accounting, and HR to automate tasks such as identifying discrepancies, processing invoices, and sorting employee records.
Available to Laserfiche Cloud users starting May 7, 2026, future integration of agents into business processes, background execution, and system monitoring capabilities will be added.

Notable Quotes & Details

Notable Data / Quotes

Karl Chan, CEO of Laserfiche, said, “The introduction of AI Agents to content management signals a change in how we handle the information lifecycle. We are moving beyond manual processes by offloading mundane work to agents that operate in a governance framework. We are letting organisations modernise operations while keeping compliance at the forefront.”
Justin Pava, Laserfiche chief product evangelist, spoke on the future of document storage, saying “the ‘where’ of document storage is not going to be as important as it used to be. With automatically-extracted metadata, AI-assisted search and the autonomous abilities of Laserfiche AI agents, you won’t have to spend time organising data, you will be able to simply act on it.”
May 7, 2026

Intended Audience

IT managers, business decision makers, process automation personnel, legal, accounting, and HR department professionals in companies using content management systems.

Using Polars Instead of Pandas: Performance Deep Dive

2026-05-12

Summary

This article compares the data processing performance of Polars and Pandas and explains why Polars outperforms Pandas on large datasets.

Key Points

Polars outperforms Pandas in data processing, especially when processing millions of rows of data.
While Pandas executes each operation sequentially, Polars is built on top of Apache Arrow, based on Rust, and optimizes query plans through parallelism and lazy evaluation.
The article uses three real-data problems from the StrataScratch coding platform to compare the solutions and performance differences of the two libraries.
Polars provides an efficient way to rank using sorting with `with_row_count` instead of the `rank` function.

Notable Quotes & Details

Notable Data / Quotes

Over the last decade, Pandas has been the foundation for data work in Python.
However, once you start working with millions of rows, the flaws start to appear: groupby operations that take several seconds, intermediate copies that consume RAM, and window functions that run as Python-level loops rather than vectorized C or Rust code.
Polars is a DataFrame library built in Rust on top of Apache Arrow.
Polars optimal solution avoids the rank function entirely. After sorting by ["total_emails", "user_id"] in descending and ascending order, respectively, the .with_row_count("activity_rank", offset=1) clause assigns sequential integers starting from 1.

Intended Audience

Data scientist, data engineer, Python developer working with large datasets

5 Useful Python Scripts for Time Series Analysis

2026-05-12

Summary

This article describes five useful Python scripts for time series data analysis.

Key Points

Five Python scripts handle common analysis tasks for time series data.
It supports CSV or Excel input and is designed to produce clean output and be easy to set up for a variety of datasets.
Includes scripts to resample irregularly spaced data and identify outliers using z-score, IQR, and rolling statistical methods.

Notable Quotes & Details

Notable Data / Quotes

z-score method flags points where the standardized value exceeds a configurable threshold (default ±3)
interquartile range (IQR) method flags points outside 1.5× the interquartile range.

Intended Audience

Time series data analysts, data scientists, and engineers

Notes: Content incomplete

I created a minimal one-file implementations (160loc) of JEPA family (ijepa, vjepa, vjepa2, cjepa) for educational purposes [P]

2026-05-12

Summary

The author has created a minimal implementation of the JEPA family (ijepa, vjepa, vjepa2, cjepa) algorithms for educational purposes.

Key Points

We developed minimal implementations of the JEPA algorithms (ijepa, vjepa, vjepa2, cjepa).
For educational purposes, we focused on understanding the nitty-gritty of the algorithm.
Each implementation is implemented in PyTorch with 160-200 lines of code, and tutorial files are also provided.

Notable Quotes & Details

Notable Data / Quotes

160-200 lines of code
https://github.com/keon/jepa

Intended Audience

Machine learning developers, students, researchers, and anyone interested in understanding the JEPA algorithm

TabPFN-3 just released: a pre-trained tabular foundation model for up to 1M rows [R][N]

2026-05-12

Summary

TabPFN-3 is a pre-trained tabular foundation model that can predict in a single forward pass on up to 1 million rows of tabular data.

Key Points

TabPFN-3 scales to 1 million rows of data, 10x more than previous versions (TabPFN-2.5, TabPFNv2), and can run on a single H100 GPU.
Inference speed has increased from 10 to 1000 times compared to previous versions.
API-exclusive 'Think Mode' increases prediction accuracy through additional learning during inference, outperforming all non-TabPFN methods in TabArena by over 200 Elo.
TabArena has a 93% win rate compared to traditional ML methods.
Includes a non-parametric search decoder supporting up to 160 classes and a regression head that produces calibrated quantile predictions in a single forward pass.
Achieve new best-in-class performance on time series, interpretability, and relational benchmarks and improve adjacency tasks.
We offer three distribution paths: API, enterprise license, and open source weighting (for research and academic evaluation purposes).

Notable Quotes & Details

Notable Data / Quotes

1M rows
H100
~8GB per million rows per estimator
10x-1000x faster inference
120x on SHAP
200 much
4-hour-tuned AutoGluon 1.5 extreme
420 much
93% win rate
160 classes
TabPFN-2.5 (Nov 2025)
TabPFNv2 (Nature, Jan 2025)
3M downloads
200+ published applications

Intended Audience

Machine learning researcher, data scientist, artificial intelligence developer

I Found a Hidden Ratio in Transformers That Predicts Geometric Stability [R]

2026-05-12

Summary

This study found that the ratio of MLP and attention spectral norm is important in predicting the geometric stability of transformer models.

Key Points

The decoder transformer model was analyzed by Lyapunov spectral analysis.
The ratio of the MLP to the attention spectral norm strongly predicts whether the model will collapse to rank-1 in the final layer.
To maintain model stability through the final layer, it is best to keep the spectral ratio between 0.5 and 2.
Related papers and GitHub repositories have been released.

Notable Quotes & Details

Notable Data / Quotes

0.5–2
https://github.com/yousef-rafat/the-1-1-rule

Intended Audience

Machine learning researchers, transformer model developers, people interested in neural network theory

Interaction Models from Thinking Machines Lab [P]

2026-05-12

Summary

Information about the interaction model developed by Thinking Machines Lab.

Notable Quotes & Details

Intended Audience

Machine learning researchers and developers

Notes: Content incomplete

IEEE Program Aims to Connect the Billions Who Are Still Offline

2026-05-12

Summary

IEEE's 'Connecting the Unconnected' program aims to develop, standardize and deploy technologies to provide connectivity to the world's unconnected population.

Key Points

More than 2 billion people, or about 30% of the world's population, still do not have access to the Internet.
IEEE Future Networks' Connecting the Unconnected (CTU) program is accelerating the development and deployment of 5G, 6G, and next-generation technologies.
CTU promotes digital inclusion by hosting a global competition and expert summit for connectivity technology innovators each year.

Notable Quotes & Details

Notable Data / Quotes

Nearly 30 percent of the global population still has no access to it. More than 2 billion people are still offline
Since 2021
Last year 245 projects from 52 countries were submitted.
US $500 to $2,500
IEEE Future Networks has created a community to bring all these initiatives working on digital connectivity together in a single platform and leverage the IEEE brand to help raise the visibility of their work.

Intended Audience

Technology developers, policy makers, social entrepreneurs, IEEE members, and the general public interested in Internet connectivity and bridging the digital divide.

Neutralizing the Gigascale Problem: How to Solve the Physical Power Paradox of Extreme AI Training Loads

2026-05-12

Summary

This is about Ampace's semi-solid-state battery solution to solve data center power supply problems caused by very large AI workloads.

Key Points

The gigascale growth of AI workloads is creating physical bottlenecks in data center power supply, due to the lack of dynamic resilience in the power chain.
Large GPU clusters generate high-frequency, sudden, synchronized, spike-pulsing loads that cause voltage sags, frequency fluctuations, and local power grid instability.
Traditional backup power sources (diesel generators, gas turbines) cannot respond to millisecond power spikes, forcing operators to overinvest in expensive infrastructure.
Ampace's PU series semi-solid-state batteries absorb millisecond power spikes with ultra-low internal resistance (DCR) and high cycle life to ensure power system stability.
At Data Center World held in Washington D.C. in 2026, Ampace revealed through discussions with Eaton that energy storage devices must evolve into 'active high-speed stabilizers' that solve the physical power paradox in the AI era.

Notable Quotes & Details

Notable Data / Quotes

Data Center World 2026 in Washington, D.C.
100 kW+
millisecond-level power spikes

Intended Audience

AI data center operators, power system designers, infrastructure engineers, energy storage technology experts

Article: Time-Series Storage: Design Choices That Shape Cost and Performance

2026-05-12

Summary

Describes various design choices for optimizing the cost and performance of time series data storage.

Key Points

Normalizing the series IDs into a separate metadata table saves approximately 42% of time series storage space.
Highly unique fields such as request ID and session token should be excluded from the series ID.
Storing series dimensions in a flexible JSON format such as PostgreSQL jsonb avoids schema migration, but requires a careful indexing policy.
Time slicing enables O(1) data expiration but creates write hotspots, and you can spread out the writes by adding a second axis (series ID).
Downsampling from 5-second resolution to 1-hour resolution reduces the number of rows by a factor of 720, improving storage efficiency.
Time series data is a sequence of measurements that tracks a history of changes over time rather than a current state.
The core data points of time series data consist of timestamps, identifiers (dimensions or tags), and values (metrics or fields).

Notable Quotes & Details

Notable Data / Quotes

forty-two percent
720 times
$50
10:00, 10:01, 10:02...
five seconds
thousands of price updates per second
O(1)

Intended Audience

Developers, data engineers, and database architects interested in designing, optimizing, and managing time series data storage.

New Exim BDAT Vulnerability Exposes GnuTLS Builds to Potential Code Execution

2026-05-12

Summary

A critical BDAT vulnerability has been discovered affecting the GnuTLS build of the Exim MTA, resulting in a potential code execution risk.

Key Points

A use-after-free vulnerability (CVE-2026-45185, CVSS score 9.8) was discovered during parsing of the BDAT message body in builds using GnuTLS in the Exim MTA.
This vulnerability occurs when a client sends a TLS close_notify warning before completing a BDAT transmission and sends the last byte as clear text on the same TCP connection, leading to heap corruption.
Only Exim builds from versions 4.97 to 4.99.2 using USE_GNUTLS=yes are affected, and this is fixed in version 4.99.3, so upgrading is recommended immediately.

Notable Quotes & Details

Notable Data / Quotes

CVE-2026-45185
CVSS score: 9.8
May 1, 2026
Exim versions from 4.97 up to and including 4.99.2
version 4.99.3
CVE-2017-16943
During TLS shutdown, Exim frees its TLS transfer buffer – but a nested BDAT receive wrapper can still process incoming bytes and end up calling ungetc(), which writes a single character (\n) into the freed region.
This sequence of events can cause Exim to write into a memory buffer that has already been freed during the TLS session teardown, leading to heap corruption.
The fix ensures that the input processing stack is cleanly reset when a TLS close notification is received during an active BDAT transfer, preventing the stale pointers from being used.

Intended Audience

Exim MTA user and system administrator, cybersecurity expert

Webinar: What the Riskiest SOC Alerts Go Unanswered - and How Radiant Security Can Help

2026-05-12

Summary

This article is about a technical webinar covering why dangerous SOC alerts aren't being processed and how Radiant Security's AI platform solves them.

Key Points

Security operations teams are overloaded with alerts, especially high-risk alerts such as WAF, DLP, OT/IoT, dark web information, and supply chain signals that often go uninvestigated.
Internal SOC teams, MSSPs, and MDRs all struggle to handle complex, specialized alerts due to lack of expertise and economic reasons.
Existing AI SOC automation platforms rely on static logic, which limits their ability to handle new threats or unfamiliar alert types.
Radiant Security's AI SOC platform handles all types of alerts by creating custom triage logic in real time instead of pre-built playbooks.
Radiant Security and German cybersecurity company Cirosec will host a technical webinar to bridge this gap on May 21, 2026.

Notable Quotes & Details

Notable Data / Quotes

WAF, DLP, OT/IoT, dark web intelligence, and supply chain signals
four to six pre-defined categories
May 21, 2026
Alert Coverage No One Else Can Triage

Intended Audience

Security experts, SOC analysts, IT decision makers, and anyone interested in cybersecurity solutions

Why Agentic AI Is Security's Next Blind Spot

2026-05-12

Summary

Agent AI is operating in organizations without proper intervention from security teams, and security experts do not properly understand this technology, creating a new security blind spot.

Key Points

In many organizations, Agent AI is already running in production environments without meaningful involvement from security teams, creating new security risks.
Security professionals lack a deep understanding of Agent AI technology, making it difficult to effectively defend and manage it.
Similar to the introduction of cloud computing, security control is lost without a technical understanding of agent AI, and if security teams do not understand the language of the technology, they are excluded from business decisions.
General-purpose coding and productivity agents such as Claude Code and GitHub Copilot, as well as vendor-built agents based on the Model Context Protocol (MCP), are already widely used, and there is a risk that agents can execute malicious instructions through attack vectors such as malicious calendar invitations.

Notable Quotes & Details

Notable Data / Quotes

genuine fluency in a technology must come before you can meaningfully defend it.
The same dynamic is playing out with AI, at a faster pace and with higher stakes.
This has played out with every major technology shift over the past two to three decades. AI will be no different.
Model Context Protocol

Intended Audience

Information security experts, AI developers, technology managers, and corporate decision makers.

Instructure Reaches Ransom Agreement with ShinyHunters to Stop 3.65TB Canvas Leak

2026-05-12

Summary

Instructure stopped a 3.65TB Canvas data leak through a ransomware settlement with ShinyHunters.

Key Points

Instructure has reached a settlement with cybercrime group ShinyHunters after facing threats of stolen data leaks following a network breach.
The company took the controversial decision to pay a ransom to avoid a data breach, with the stolen data returned and a digital confirmation of data destruction.
ShinyHunters stole 3.65 TB of data, affecting approximately 9,000 organizations, and a second unauthorized activity was detected on May 7, 2026, where the Canvas login portal was tampered with.
Attackers exploited a vulnerability related to 'support tickets' in the Free-for-Teacher environment to steal 275 million user names, email addresses, course names, registration information, and message history.
Instructure emphasized that course content, submissions, and credentials were not compromised, and temporarily closed the Free-For-Teacher account.

Notable Quotes & Details

Notable Data / Quotes

agreement
concerns about the potential publication of data.
While there is never complete certainty when dealing with cyber criminals, we believe it was important to take every step within our control to give customers additional peace of mind, to the extent possible
3.65TB
9,000 organizations
May 7, 2026
330 institutions
May 12, 2026
275 million records
The exfiltrated data provides threat actors enough personal context to conduct targeted phishing campaigns against staff, students, and parents alike

Intended Audience

Educational technology users, cybersecurity professionals, and school personnel using Instructure or Canvas.

OpenAI Launches Daybreak for AI-Powered Vulnerability Detection and Patch Validation

2026-05-12

Summary

OpenAI has launched Daybreak, a new cybersecurity initiative that combines AI models with Codex Security to help enterprises detect and patch vulnerabilities.

Key Points

OpenAI Daybreak leverages AI models and Codex Security to proactively identify and patch enterprise vulnerabilities.
This initiative strengthens software resilience by integrating secure code reviews, threat modeling, patch verification, and more into the development process.
Daybreak is based on three models: GPT-5.5, GPT-5.5 with Trusted Access for Cyber, and GPT-5.5-Cyber.
Advances in AI technology are accelerating the speed of vulnerability discovery, causing issues with the patching process and ‘triage fatigue’.

Notable Quotes & Details

Notable Data / Quotes

Earlier this March
the 90 day disclosure policy is dead
Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler

Intended Audience

Cybersecurity expert, software developer, enterprise IT manager, AI and security technology investor

PreviousDaily Briefing

NextDaily Briefing