Pablo Villegas

Valiente como un Jedi sin sable. Tech, mkt & economics you could find over here. PJ fan 4ever

[Spotify @Spotify]

💡 Prompt of the Week 💡

Create a Prompted Playlist with the below and see what audiobooks you get:

Give me your best thrillers. I don't want to see the twists coming. I hate plot-holes. I need something that will keep me guessing at the end of every chapter. And I want the payoff at the end to be mind-blowing!

[DiscussingFilm @DiscussingFilm]

‘SHŌGUN’ Season 2 has wrapped filming.

• Set 10 years after the first season

• Hiroyuki Sanada and Cosmo Jarvis will reprise their roles

[Pablo Villegas @Pablo_Ved]

Ajá claro, como no.

[Engadget @engadget]

Windows laptops aren't known for having excellent battery life, but a lot of that is because of unoptimized settings. Change these for big improvements. https://t.co/MkdEmnA8pj

[CyberSignal | Cybersecurity & AI News @XQOPTRX]

🚨 [AI-ERA PHISHING EVASION] — ATTACKERS ARE USING INVISIBLE UNICODE CHARACTERS TO HIDE WORDS INSIDE MASS PHISHING EMAILS

Microsoft says attackers repurposed “ASCII Smuggling,” a technique known from AI prompt-injection research, to evade email-content analysis at MULTI-MILLION-MESSAGE DAILY SCALE.

CyberSignal Priority: 🔴 VERY HIGH

📅 September 3, 2026
🏢 Microsoft Security Research
🎣 Threat: High-volume phishing
🧠 Technique: ASCII Smuggling / Unicode Tags
📊 Peak activity: Millions of messages per day
🏷️ Phishing · AI Security · Unicode · Email Security · Evasion

### 🔎 What happened

Microsoft discovered attackers abusing invisible characters from the Unicode Tags block:

U+E0000 → U+E007F

These characters can exist inside a message while remaining invisible to the person reading it.

Attackers inserted them inside words used in financial phishing lures.

Human sees:

FUNDING

Detection engine may effectively process something closer to:

FUN[invisible characters]DING

The objective:

BREAK SIMPLE TEXT ANALYSIS WITHOUT CHANGING WHAT THE VICTIM SEES.

Microsoft observed activity increase sharply beginning in February 2026 and reach MULTI-MILLION-MESSAGE daily volumes.

### ⚔️ Attack chain

Mass phishing campaign
↓
financial/social-engineering lure
↓
invisible Unicode Tags inserted inside keywords
↓
email looks completely normal to victim
↓
text processing sees altered character sequence
↓
keyword/signature detection may be weakened
↓
phishing message reaches target

### 🤖 Why AI matters

ASCII Smuggling became especially notable through:

AI PROMPT-INJECTION RESEARCH.

The original concept:

Invisible text
↓
human cannot see instruction
↓
AI processes underlying characters
↓
hidden prompt influences model.

Attackers are now applying the same representation problem in another direction:

NOT HIDING TEXT FROM HUMANS SO AI CAN READ IT.

HIDING TEXT FROM DEFENSIVE SOFTWARE WHILE HUMANS CAN STILL UNDERSTAND IT.

### 🎯 What is affected

Potentially vulnerable processing layers include:

email gateways
spam filters
content-inspection engines
DLP systems
AI email assistants
text-normalization pipelines
custom phishing detectors.

Any system that interprets invisible Unicode differently from the user may expose an evasion gap.

### 🧠 Why this matters

AI security and conventional cybersecurity are beginning to converge.

Techniques researched for:

LLM PROMPT INJECTION

can potentially be repurposed for:

PHISHING EVASION
CONTENT FILTER BYPASS
SECURITY TOOL EVASION.

The attack surface is no longer only:

WHAT TEXT SAYS.

It is also:

HOW SOFTWARE REPRESENTS THAT TEXT.

### ⚠️ Important caveat

ASCII Smuggling does NOT automatically bypass modern email-security platforms.

Microsoft says Defender for Office 365 detected more than 99% of observed messages through other defensive signals including:

sender reputation
IP reputation
domain/URL reputation
machine-learning phishing detection
brand impersonation detection
email-authentication signals.

That is exactly why layered detection matters.

### 🛡️ Defender action

Security teams should verify that mail-processing systems:

normalize Unicode before analysis
detect Unicode Tag characters
compare rendered and raw message content
use URL/domain reputation
use behavioral and ML detection
do not rely solely on keyword matching.

Specifically inspect handling of:

U+E0000–U+E007F.

### 🧠 CyberSignal insight

Every new representation of text creates another place for attackers to hide.

THE HUMAN MAY SEE ONE MESSAGE.

THE SECURITY ENGINE MAY SEE ANOTHER.

AND THE AI MODEL MAY SEE SOMETHING DIFFERENT AGAIN.

Sources: Microsoft Security Research

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚨 [AI-ASSISTED CYBERATTACKS] — ATTACKERS ARE USING COMMERCIAL LLMs TO TROUBLESHOOT REAL-WORLD NETWORK INTRUSIONS

Palo Alto Networks Unit 42 found attackers targeting government, transportation, water and financial organizations using AI to repeatedly rewrite scripts, overcome failed commands and support intrusion operations.

CyberSignal Priority: 🔴 VERY HIGH

📅 September 3, 2026
🔬 Palo Alto Networks Unit 42
🌎 Mexico · Ecuador · Brazil
👤 CL-CRI-1131 · CL-CRI-1163
🤖 Commercial LLMs · NextChat
🏷️ AI-Assisted Attacks · Espionage · RAT · Exfiltration

This is NOT:

“AI could theoretically help hackers.”

Researchers found evidence of AI being integrated into:

REAL ATTACK OPERATIONS.

### 🔎 What happened

Unit 42 disclosed two separate intrusion clusters operating against Latin American organizations.

CL-CRI-1131 targeted organizations including:

government ministries
transportation
municipal water infrastructure
organizations in Mexico and Ecuador.

CL-CRI-1163 targeted:

BRAZILIAN FINANCIAL ORGANIZATIONS.

Both campaigns displayed evidence consistent with attackers incorporating commercial LLMs into their workflows.

### ⚔️ Attack chain — first cluster

System compromised
↓
attacker attempts SAM / NTDS.dit extraction
↓
commands repeatedly fail
↓
scripts repeatedly rewritten
↓
new variants executed
↓
shadow copies created
↓
credentials/sensitive information collected
↓
attacker infrastructure
↓
exfiltration

Researchers observed multiple successive script versions apparently created to work around technical failures.

### 🤖 AI infrastructure

Unit 42 discovered infrastructure hosting:

NEXTCHAT

on TCP port:

3000.

NextChat provides a single interface through which operators can interact with multiple LLM providers.

Researchers assess that commercial language models were likely being used to:

debug scripts
generate workaround commands
modify tooling
solve execution errors
accelerate offensive development.

### 🇧🇷 Brazilian campaign

The second cluster used:

resume-themed phishing
↓
remote-access malware
↓
Go-based SOCKS5 tunneling
↓
failed deployment
↓
tool modified
↓
retried
↓
modified again.

Unit 42 observed versions:

1 → 9

of one tunneling utility being attempted within approximately:

TWO HOURS.

Exposed attacker infrastructure also contained scripts with names such as:

exploit_creative.py
exploit_careful.py
rce_focused.py

Researchers said the iterative development pattern was consistent with language-model-assisted coding.

### 🎯 What was targeted

Observed sectors included:

government
transportation
municipal water
financial services.

The important shift:

AI-assisted offensive workflows are no longer confined to:

LABS
CTFs
RED-TEAM DEMOS.

They are appearing inside:

REAL INTRUSIONS.

### 🧠 Why this matters

The most important AI advantage may not be:

AUTONOMOUS HACKING.

It may simply be:

REMOVING FRICTION.

An attacker gets an error.

Previously:

research
search forums
write new code
debug manually.

Now:

paste error into LLM
↓
receive modified script
↓
run
↓
paste next error
↓
repeat.

AI can dramatically accelerate the attacker's learning loop.

### ⚠️ Important caveat

Unit 42 did NOT conclude that autonomous AI independently conducted these campaigns.

Humans remained involved.

The evidence supports:

AI-ASSISTED OPERATIONS.

Not:

FULLY AUTONOMOUS CYBERATTACKS.

The attackers also demonstrated poor operational security, exposing infrastructure and tooling that helped researchers reconstruct their activity.

### 🛡️ Defender action

Do not waste time trying to detect whether source code “looks AI-generated.”

Detect the attack behavior:

repeated script variations
unexpected NextChat exposure
dynamic DNS
SOCKS5 tunneling
Certutil downloads
shadow-copy creation
SAM / NTDS.dit access
credential dumping
open staging directories.

### 🧠 CyberSignal insight

AI does not need to turn an average attacker into an autonomous super-hacker.

It only needs to make:

EVERY FAILED COMMAND EASIER TO FIX.

That alone changes offensive economics.

Sources: Palo Alto Networks Unit 42

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚨 [JUST IN — MICROSOFT AI] — MICROSOFT LAUNCHES MAI-TRANSCRIBE-2, CLAIMING ONE HOUR OF AUDIO CAN BE PROCESSED IN ABOUT 10 SECONDS

The new speech-recognition model supports 60 languages, speaker diarization and word-level timestamps — with limited-time API pricing of $0.10 PER AUDIO HOUR.

CyberSignal AI Priority: 🔴 VERY HIGH

📅 September 3, 2026
🏢 Microsoft AI
🧠 Model: MAI-Transcribe-2
🌍 60 languages
⚡ 1 hour audio → ~10 sec inference
💰 $0.10/audio hour limited-time price
🏷️ Voice AI · Speech Recognition · AI Agents

Microsoft just pushed down one of the major costs behind large-scale:

VOICE AI.

### 🔎 What happened

Microsoft AI officially launched:

MAI-TRANSCRIBE-2.

The model combines:

speech recognition
speaker diarization
word-level timestamps
multilingual transcription
domain/context biasing.

All inside one speech-processing system.

### ⚙️ Pipeline

Audio
↓
speech recognition
↓
language/accent handling
↓
speaker identification
↓
word-level timestamps
↓
domain terminology processing
↓
structured transcript

### 📊 Microsoft's reported results

Microsoft reports support for:

60 LANGUAGES.

Average Word Error Rate across its top 25 evaluated languages on FLEURS:

3.4%.

Microsoft also reports approximately:

2% WER

on Artificial Analysis.

The company says MAI-Transcribe-2 performs competitively against systems including:

Whisper-Large-V3
GPT-Transcribe
ScribeV2
Gemini transcription systems

under its reported benchmark conditions.

### ⚡ Speed

Microsoft reports:

1 HOUR OF AUDIO
↓
≈10 SECONDS OF MODEL INFERENCE.

Its previous MAI-Transcribe-1.5 reportedly required roughly:

20 SECONDS.

That represents approximately:

2× FASTER

under Microsoft's comparison setup.

### 💰 Price

Limited-time API price:

$0.10 PER AUDIO HOUR.

That is particularly important.

At that price:

1,000 hours = ~$100
10,000 hours = ~$1,000
100,000 hours = ~$10,000

before additional infrastructure/application costs.

Speech processing is becoming cheap enough to operate at enormous scale.

### 🎯 Where this matters

call centers
meeting transcription
voice agents
medical documentation
podcasts
subtitles
compliance
customer intelligence
searchable video
accessibility
enterprise archives.

### 🧠 Why this matters

Voice AI has several layers:

LISTEN
↓
UNDERSTAND
↓
REASON
↓
RESPOND
↓
ACT.

If the first layer is inaccurate, everything downstream becomes unreliable.

Better and cheaper transcription therefore improves much more than:

SUBTITLES.

It becomes infrastructure for:

VOICE AGENTS.

### ⚠️ Important caveat

Microsoft describes MAI-Transcribe-2 as the:

“fastest, most accurate and cheapest”

speech-recognition model.

That is Microsoft's own product claim.

Real performance varies according to:

language
accent
background noise
speaker overlap
hardware
recording quality
specialized terminology.

Independent testing remains necessary.

Also:

$0.10/hour

is explicitly LIMITED-TIME pricing.

### 🧠 CyberSignal insight

The biggest transformation in voice AI may not come from machines learning to:

SPEAK BETTER.

It may come from making machines cheap enough to:

LISTEN AT PLANETARY SCALE.

Sources: Microsoft AI

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚨 [JUST IN — GOOGLE AI] — GOOGLE LAUNCHES GMAIL LIVE, DOCS LIVE AND KEEP LIVE, TURNING WORKSPACE INTO A CONVERSATIONAL VOICE INTERFACE

Gemini can search your inbox through conversation, transform spoken brainstorming into documents and automatically organize voice “brain dumps” into structured notes.

CyberSignal AI Priority: 🔴 VERY HIGH

📅 September 3, 2026
🏢 Google · Gemini · Workspace
🎙️ Gmail Live · Docs Live · Keep Live
🏷️ Voice AI · Productivity · AI Agents · Gemini

Google is beginning to replace:

CLICK
+
TYPE

with:

TALK.

### 🔎 What happened

Google officially announced three Gemini-powered voice capabilities:

GMAIL LIVE
DOCS LIVE
KEEP LIVE.

They were previewed earlier, but their rollout announcement is dated:

SEPTEMBER 3, 2026.

### 📧 Gmail Live

Instead of manually searching through email:

User asks a question
↓
Gemini searches relevant messages
↓
context is retrieved
↓
information is synthesized
↓
answer returned conversationally.

Example:

“WHAT GATE IS MY FLIGHT LEAVING FROM?”

Rather than manually finding the confirmation email, Gemini can search the inbox and provide the answer conversationally.

### 📄 Docs Live

Docs Live moves beyond dictation.

The user can:

TALK THROUGH AN IDEA.

Gemini can then help:

organize the ideas
structure sections
draft content
refine wording
develop the document.

With appropriate permission, Docs can also incorporate relevant context from:

Gmail
Drive
Chat
the web.

So the workflow becomes:

UNSTRUCTURED HUMAN THOUGHT
↓
VOICE
↓
GEMINI
↓
CONTEXT RETRIEVAL
↓
STRUCTURED DOCUMENT.

### 📝 Keep Live

Keep Live is designed for what Google essentially describes as:

A BRAIN DUMP.

User speaks ideas
↓
Gemini extracts information
↓
ideas organized
↓
tasks separated
↓
notes structured
↓
actionable information preserved.

Instead of forcing the human to structure thoughts before entering them:

THE AI STRUCTURES THEM AFTERWARD.

### 🧠 Why this matters

For decades, productivity software has forced humans to translate ideas into:

MENUS
FILES
BUTTONS
FOLDERS
FORMATTING.

AI agents are beginning to reverse that relationship.

The human communicates naturally.

The software decides:

where to search
what information matters
how to organize it
what application action is required.

That is the transition from:

SOFTWARE TOOL

toward:

SOFTWARE AGENT.

### 🎯 What changes

Traditional Gmail:

SEARCH → OPEN → READ → SYNTHESIZE.

Gmail Live:

ASK.

Traditional Docs:

PLAN → TYPE → FORMAT → REWRITE.

Docs Live:

TALK → AI STRUCTURES.

Traditional Keep:

WRITE → ORGANIZE → LABEL.

Keep Live:

SPEAK → AI ORGANIZES.

### 🔐 Security implication

More capable assistants also create larger permission boundaries.

An assistant connected simultaneously to:

email
documents
chat
cloud files
web search

can potentially access a large amount of sensitive organizational context.

Enterprise controls therefore become increasingly important:

least privilege
data boundaries
connector permissions
audit logs
DLP
prompt-injection defenses
sensitive-data classification.

### ⚠️ Important caveat

These capabilities do NOT mean Gemini autonomously has unrestricted access to all Workspace data.

Access depends on:

user permissions
enabled services
organization settings
available integrations.

Organizations should evaluate these controls before enabling broad agentic workflows.

### 🧠 CyberSignal insight

The next major UI revolution may not be another:

APP.

It may be the disappearance of the interface itself.

EMAIL
DOCUMENTS
NOTES
FILES

could increasingly become one conversational layer where the user simply says:

WHAT THEY WANT DONE.

Sources: Google Workspace · Google Gemini

[Mariano Amartino @amartino]

El CEO de AMC Theaters al CEO de Robinhood explicando porque esta en contra de la tokenización sintética de sus acciones... y liteeralmente le dice que va a buscar como iniciar acciones contra $HOOD por falsificar acciones de $AMC

Hay algo clave que no me canso de repetir:
- NO compras una accion al comprar un token que emite un tercero si no es nativo.
- AUN cuando te arriesgues, como compras algo que NO es una acción real, no estas comprando los derechos que vienen con las mismas (desde derecho a voto a otras cosas)
- Si un dia el emisor del token (NO NATIVO) desaparece, desaparece tu plata :)

[Marcela Mexia 🖤 @MarcelaMexia]

Para los CEOs que todavía creen que marketing es “el área que hace anuncios”:

Sin marketing estratégico, tu empresa queda ciega, sorda y muda.

https://t.co/nwMrLxiSE9

[ZDNET @ZDNET]

4 AI development skills you need, according to Andrew Ng - and what experts say he's missing https://t.co/u2St4hIa3D

[El Grupo Informático @EGInformatico]

La IA de ChatGPT ya permite crear stickers para WhatsApp: te explicamos cómo https://t.co/nfxQPIEIUU

[Techmeme @Techmeme]

WhatsApp upgrades its two-step verification, letting users replace the six-digit PIN with a longer alphanumeric password, and adds support for multiple passkeys (@aiishamalik1 / TechCrunch)

(Visit Techmeme dot com for the link and full context!)

[Pablo Villegas @Pablo_Ved]

@airtminc B y C

[Bale(gas投资人) @BaleYoba]

美国为什么这个时候人为制造crypto 的牛市？

原因无它，就两个字:化债。

美债利率攀升，说明美债没人买了，川普智囊团给出的解决思路就是—短期，政府买，长期币圈买。币圈如何买？用美元稳定币买，因为稳定币的储备资产必须买美债。

这是一个顶级阳谋:
美股在全球率先链上代币化，美股作为全球优质资产，7/24交易带来的全球交易增量，必然会带来链上繁荣，链上繁荣会带动加密二哥 ethereum:native 的持续向好，链上和股票会形成螺旋上升互动，稳定币发行总量逐步拉升，美债购买力持续加强。

这是国运级别的较量，确保自己要在车上，才能继续看历史剧。

[Healthy Pockets @healthy_pockets]

“BITCOIN ESTA MUERTO”

Los holders de BTC esta semana 😂😂

[ether.fi @ether_fi]

Two products millions of DeFi users already rely on.

One thing they've never done together.

Next week 👀

[Pablo Villegas @Pablo_Ved]

Wauw!!! Qué sorpresa. Nadie se lo veía venir #not

[Accionero @AccioneroCom]

ÚLTIMA HORA: 🇺🇸 Presidente Trump dice que EE. UU. está considerando comprar cantidades "considerables" de Bitcoin y otras criptomonedas.

[Duolingo Español @duolingoespanol]

Wisin abrió la Universidad del Perreo y yo pensando que no tenía competencia en el sector educativo.

[Bessy Granja Barriga @bessygranjaOK]

🙏🇨🇴 Oremos por Colombia.
Señor, extiende tu mano sobre cada familia que hoy vive momentos de angustia tras el terremoto. Dale fortaleza a este pueblo hermano y permite que no haya pérdidas humanas que lamentar. 🙏
Amén.
Mi solidaridad con ustedes hermanos colombianos. 💔

[Christine @chris_suta]

Hoy colocaron en X finalmente el botón para ver tus impresiones generales y las que generan los verificados, por si estaban con el pendiente 🫠

[Pablo Villegas @Pablo_Ved]

Todos los políticos hacen lo mismo . Nadie en esta red social , es su Target sino los barrios pobres porque allá se encuentra la mayoría del voto duro. Los que ponen presidentes y autoridades. En conclusión 1 verde por tu voto y no importa plan de gobernanza alguno.

[Carloncho🇪🇨 @carlozuri]

Me quiero cortar los dedos de pensar que vote por esta gentusa

[ether.fi @ether_fi]

not letting go of that 3% cashback

[Pablo Villegas @Pablo_Ved]

You know.... You know where you are with ...