Quote for the day:
"It is during our darkest moments that we must focus to see the light." -- Aristotle Onassis
Machine unlearning gets a practical privacy upgrade
Machine unlearning, which refers to strategies for removing the influence of
specific training data from a model, has emerged to fill the gap. But until now,
most approaches have either been slow and costly or fast but lacking formal
guarantees. A new framework called Efficient Unlearning with Privacy Guarantees
(EUPG) tries to solve both problems at once. Developed by researchers at the
Universitat Rovira i Virgili in Catalonia, EUPG offers a practical way to forget
data in machine learning models with provable privacy protections and a lower
computational cost. Rather than wait for a deletion request and then scramble to
rework a model, EUPG starts by preparing the model for unlearning from the
beginning. The idea is to first train on a version of the dataset that has been
transformed using a formal privacy model, either k-anonymity or differential
privacy. This “privacy-protected” model doesn’t memorize individual records, but
still captures useful patterns. ... The researchers acknowledge that extending
EUPG to large language models and other foundation models will require further
work, especially given the scale of the data and the complexity of the
architectures involved. They suggest that for such systems, it may be more
practical to apply privacy models directly to the model parameters during
training, rather than to the data beforehand.Emerging Cloaking-as-a-Service Offerings are Changing Phishing Landscape
Cloaking-as-a-service offerings – increasingly powered by AI – are “quietly
reshaping how phishing and fraud infrastructure operates, even if it hasn’t
yet hit mainstream headlines,” SlashNext’s Research Team wrote Thursday. “In
recent years, threat actors have begun leveraging the same advanced
traffic-filtering tools once used in shady online advertising, using
artificial intelligence and clever scripting to hide their malicious payloads
from security scanners and show them only to intended victims.” ... The newer
cloaking services offer advanced detection evasion techniques, such as
JavaScript fingerprinting, device and network profiling, machine learning
analysis and dynamic content swapping, and put them into user-friendly
platforms that hackers and anyone else can subscribe to, SlashNext researchers
wrote. “Cybercriminals are effectively treating their web infrastructure with
the same sophistication as their malware or phishing emails, investing in
AI-driven traffic filtering to protect their scams,” they wrote. “It’s an arms
race where cloaking services help attackers control who sees what online,
masking malicious activity and tailoring content per visitor in real time.
This increases the effectiveness of phishing sites, fraudulent downloads,
affiliate fraud schemes and spam campaigns, which can stay live longer and
snare more victims before being detected.”You’re Not Imagining It: AI Is Already Taking Tech Jobs
It’s difficult to pinpoint the exact motivation behind job cuts at any given
company. The overall economic environment could also be a factor, marked by
uncertainties heightened by President Donald Trump’s erratic tariff plans.
Many companies also became bloated during the pandemic, and recent layoffs
could still be trying to correct for overhiring. According to one report
released earlier this month by the executive coaching firm Challenger, Gray
and Christmas, AI may be more of a scapegoat than a true culprit for layoffs:
Of more than 286,000 planned layoffs this year, only 20,000 were related to
automation, and of those, only 75 were explicitly attributed to artificial
intelligence, the firm found. Plus, it’s challenging to measure productivity
gains caused by AI, said Stanford’s Chen, because while not every employee may
have AI tools officially at their disposal at work, they do have unauthorized
consumer versions that they may be using for their jobs. While the technology
is beginning to take a toll on developers in the tech industry, it’s actually
“modestly” created more demand for engineers outside of tech, said Chen.
That’s because other sectors, like manufacturing, finance, and healthcare, are
adopting AI tools for the first time, so they are adding engineers to their
ranks in larger numbers than before, according to her research.The architecture of culture: People strategy in the hospitality industry
Rewards and recognitions are the visible tip of the iceberg, but culture sits
below the surface. And if there’s one thing that I’ve learned over the years,
it’s that culture only sticks when it’s felt, not just said. Not once a year,
but every single day. Hilton’s consistent recognition as a Great Place to
Work® globally and in India stems from our unwavering support and commitment
to helping people thrive, both personally and professionally. ... What has
sustained our culture through this growth is a focus on the everyday. It is
not big initiatives alone that shape how people feel at work, but the smaller,
consistent actions that build trust over time. Whether it is how a team huddle
is run, how feedback is received, or how farewells are handled, we treat each
moment as an opportunity to reinforce care and connection. ... Equally vital
is cultivating culturally agile, people-first leaders. South Asia’s diversity,
across language, faith, generation, and socio-economic background, demands
leadership that is both empathetic and inclusive. We’re working to embed this
cultural intelligence across the employee journey, from hiring and onboarding
to ongoing development and performance conversations, so that every team
member feels genuinely seen and supported.
Capturing carbon - Is DAC a perfect match for data centers?
The commercialization of DAC, however, faces several significant challenges. One
primary obstacle is navigating different compliance requirements across
jurisdictions. Certification standards vary significantly between regions like
Canada, the UK, and Europe, necessitating differing approaches in each
jurisdiction. However, while requiring adjustments, Chadwick argues that these
differences are not insurmountable and are merely part of the scaling process.
Beyond regulatory and deployment concerns, achieving cost reductions is a
significant challenge. DAC remains highly expensive, costing an average of $680
per ton to produce in 2024, according to Supercritical, a carbon removal
marketplace. In comparison, Biochar has an average price of $165 per ton, and
enhanced rock weathering has an average price of $310 per ton. In addition, the
complexity of DAC means up-front costs are much higher than those of alternative
forms of carbon removal. An average DAC unit comprises air-intake manifolds,
absorption and desorption towers, liquid-handling tanks, and bespoke
site-specific engineering. DAC also requires significant amounts of power to
operate. Recent studies have shown that the energy consumption of fans in DAC
plants can range from 300 to 900 kWh per ton of CO2 captures, which represents
between 20 - 40 percent of total DAC system energy usage.
Rethinking Risk: The Role of Selective Retrieval in Data Lake Strategies
Selective retrieval works because it bridges the gap between data engineering
complexity and security usability. It gives teams options without asking them to
reinvent the wheel. It also avoids the need to bring in external tools during a
breach investigation, which can introduce latency, complexity, or worse, gaps in
the chain of custody. What’s compelling about this approach is that it doesn’t
require businesses to abandon existing tools or re-architect their
infrastructure. ... This model is especially relevant for mid-size IT teams who
want to cover their audit requirements, but don’t have a 24/7 security
operations center. It’s also useful in regulated sectors such as healthcare,
financial services, and manufacturing where data retention isn’t optional, but
real-time analysis for everything isn’t practical. ... Data volumes are
continuing to rise. As organizations face high costs and fatigue, those that
thrive will be the ones that treat storage and retrieval as distinct functions.
The ability to preserve signal without incurring ongoing noise costs will become
a critical enabler for everything from insider threat detection to regulatory
compliance. Selective retrieval isn’t just about saving money. It’s about
regaining control over data sprawl, aligning IT resources with actual risk, and
giving teams the tools they need to ask, and answer, better questions.
Manufactured Madness: How To Protect Yourself From Insane AIs
The core of the problem lies in a well-intentioned but flawed premise: that we
can and should micromanage an AI’s output to prevent any undesirable outcomes.
These “guardrails” are complex sets of rules and filters designed to stop the
model from generating hateful, biased, dangerous, or factually incorrect
information. In theory, this is a laudable goal. In practice, it has created a
generation of AIs that prioritize avoiding offense over providing truth. ...
Compounding the problem of forced outcomes is a crisis of quality. The data
these models are trained on is becoming increasingly polluted. In the early
days, models were trained on a vast, curated slice of the pre-AI internet. But
now, as AI-generated content inundates every corner of the web, new models are
being trained on the output of their predecessors. ... Given this landscape, the
burden of intellectual safety now falls squarely on the user. We can no longer
afford to treat AI-generated text with passive acceptance. We must become
active, critical consumers of its output. Protecting yourself requires a new
kind of digital literacy. First and foremost: Trust, but verify. Always. Never
take a factual claim from an AI at face value. Whether it’s a historical date, a
scientific fact, a legal citation, or a news summary, treat it as an unconfirmed
rumor until you have checked it against a primary source.
6 Key Lessons for Businesses that Collect and Use Consumer Data
Ensure your privacy notice properly discloses consumer rights, including the right to access, correct, and delete personal data stored and collected by businesses, and the right to opt-out of the sale of personal data and targeted advertising. Mechanisms for exercising those rights must work properly, with a process in place to ensure a timely response to consumer requests. ... Another issue that the Connecticut AG raised was that the privacy notice was “largely unreadable.” While privacy notices address legal rights and obligations, you should avoid using excessive legal jargon to the extent possible and use clear, simple language to notify consumers about their rights and the mechanisms for exercising those rights. In addition, be as succinct as possible to help consumers locate the information they need to understand and exercise applicable rights. ... The AG provided guidance that under the CTDPA, if a business uses cookie banners to permit a consumer to opt-out of some data processing, such as targeted advertising, the consumer must be provided with a symmetrical choice. In other words, it has to be as clear and as easy for the consumer to opt out of such use of their personal data as it would be to opt in. This includes making the options to accept all cookies and to reject all cookies visible on the screen at the same time and in the same color, font, and size.How agentic AI Is reshaping execution across BFSI
Several BFSI firms are already deploying agentic models within targeted areas of
their operations. The results are visible in micro-interventions that improve
process flow and reduce manual load. Autonomous financial advisors, powered by
agentic logic, are now capable of not just reacting to user input, but
proactively monitoring markets, assessing customer portfolios, and recommending
real-time changes.. In parallel, agentic systems are transforming customer
service by acting as intelligent finance assistants, guiding users through
complex processes such as mortgage applications or claims filing. ... For
Agentic AI to succeed, it must be integrated into operational strategy. This
begins by identifying workflows where progress depends on repetitive human
actions that follow predictable logic. These are often approval chains,
verifications, task handoffs, and follow-ups. Once identified, clear rules need
to be defined. What conditions trigger an action? When is escalation required?
What qualifies as a closed loop? The strength of an agentic system lies in its
ability to act with precision, but that depends on well-designed logic and
relevant signals. Data access is equally important. Agentic AI systems require
context. That means drawing from activity history, behavioural cues, workflow
states and timing patterns.
Open Source Is Too Important To Dilute
The unfortunate truth is that these criteria don’t apply in every use case.
We’ve seen vendors build traction with a truly open project. Then, worried about
monetization or competition, they relicense it under a “source-available” model
with restrictions, like “no commercial use” or “only if you’re not a
competitor.” But that’s not how open source works. Software today is deeply
interconnected. Every project — no matter how small or isolated — relies on
dependencies, which rely on other dependencies, all the way down the chain. A
license that restricts one link in that chain can break the whole thing. ...
Forks are how the OSS community defends itself. When HashiCorp relicensed
Terraform under the Business Source License (BSL) — blocking competitors from
building on the tooling — the community launched OpenTofu, a fork under an
OSI-approved license, backed by major contributors and vendors. Redis’
transition away from Berkeley Software Distribution (BSD) to a proprietary
license was a business decision. But it left a hole — and the community forked
it. That fork became Valkey, a continuation of the project stewarded by the
people and platforms who relied on it most. ... The open source brand took
decades to build. It’s one of the most successful, trusted ideas in software
history. But it’s only trustworthy because it means something.
No comments:
Post a Comment