Today’s three AI updates
are not about
whose model scores higher.
First, an AI agent originally designed to help enterprises find vulnerabilities
was reportedly used in bank attacks.
Second, a rapidly growing AI surveillance company
is now facing privacy backlash and workforce reductions.
Third, an Nvidia-supported AI data center company
that hoped for a major IPO
has withdrawn its listing plan.
These three cases highlight the real-world challenges after AI’s initial launch:
Can the tools be controlled?
Is society willing to accept them?
Can their business models deliver real returns?
1. ARTEX—Originally a security testing tool, now linked to South Korean bank attacks
The first case involves:
ARTEX.
It is an open-source AI Agentic Penetration Testing Tool
that was released just this year.
Its original purpose was to help
enterprises and security teams automate some penetration testing tasks,
to find system weaknesses.
Simply put, it was designed to:
use AI to help defenders find vulnerabilities.
However, CrowdStrike’s October 7 Threat Intelligence report revealed
that in investigating attacks on South Korean financial institutions,
data such as:
ARTEX configuration files,
Claude code session histories,
and Claude memory files
were found.
CrowdStrike believes attackers integrated ARTEX along with multiple large models
into their offensive workflows.
This was more than just “AI helped write some code”
CrowdStrike says the attacks lasted from:
late September
through early October,
targeting financial institutions in South Korea,
with actual data exfiltration occurring.
Reuters further reported that at least:
nine South Korean banks
have publicly confirmed or reported attacks,
and South Korean police have launched investigations.
This goes beyond the usual:
"Researchers testing whether AI can hack” scenarios.
The AI agent has been tied to real attack operations.
ARTEX itself is not a new large language model
This is an important point to clarify.
ARTEX is not another ChatGPT.
It is more like a:
Security Agent Framework
that links together:
targets,
tools,
steps,
and commands,
to form an automated workflow,
which then calls other models for assistance and execution.
In the attack environment CrowdStrike examined,
models like:
DeepSeek and
GLM were used.
Reuters noted that ARTEX was designed to connect to external models such as:
ChatGPT,
Claude,
and DeepSeek.
The key takeaway is:
AI’s dangers don’t stem only from the models themselves but from the combination of models, tools, permissions, and workflows.
The developer decides to make ARTEX closed source
The ARTEX developer known by the GitHub handle:
Autumn-27
announced after this incident that,
due to abuse of the tool,
ARTEX will:
stop public updates,
and switch to
closed source.
No new versions will be publicly released,
and maintenance support will cease.
Reuters confirmed that the original GitHub page has been taken down.
The developer stated that ARTEX’s initial goal was
to assist organizations in security testing and
they opposed any illegal use.
But shutting the repository doesn’t make already leaked versions disappear
This is today’s crucial boundary.
Once an open-source tool has been:
downloaded,
forked,
mirrored,
and copied,
shutting down the original repository
does not wipe out all existing copies.
So,
closed source
can prevent further official public distribution,
but it can’t reverse what’s already out there.
This is the hardest issue with dual-use AI tools.
The same capabilities can:
help defenders find vulnerabilities,
or help attackers exploit them faster.
True risk control cannot rely on:
“Users, please don’t break the law.”
What practical warnings does this hold for companies?
If your company begins using AI for:
vulnerability scanning,
penetration testing,
or network testing,
don’t assume that labeling it “security”
means it deserves:
full internal network access,
production credentials,
or complete internet access.
More reasonable measures are to:
restrict tested targets,
limit network egress,
log all tool calls, credential use, and outbound connections,
and require human approval before high-risk actions.
This is not because AI security tools shouldn’t be used.
On the contrary,
as they grow more capable,
boundaries must be even clearer than before.
2. Flock Safety reportedly plans 18% layoffs: AI surveillance faces public acceptance challenges
The second case involves:
Flock Safety.
If you’ve been following SasaDaily lately,
you may be familiar with this company.
Flock has deployed extensively in the US:
AI-powered license plate readers
and other camera systems.
Law enforcement uses them to search for:
specific license plates,
vehicle types,
and movement histories.
It’s become one of the largest public safety technology platforms in the US.
But today's news is not about a new abuse case.
It’s about:
growing public backlash directly impacting the company.
Reuters: Flock plans to lay off about 270 employees
According to Reuters citing sources familiar with the plan,
Flock is preparing to cut roughly:
18%
of its workforce,
which is about:
270 employees.
The company currently employs around:
1,500 people.
The layoffs are expected to occur by the end of October,
following an earlier Voluntary Separation Program.
Reuters also reported that Flock declined to comment on the layoffs,
so the correct phrasing is:
reportedly planning layoffs,
not officially announced.
The company does have a market
Quite the opposite.
Reuters reported that Flock has deployed about
120,000 AI-powered cameras
across 49 US states,
with customers including
more than 4,800 law enforcement agencies
and nearly
1,000 businesses.
The company has raised over $950 million.
This is not a case of:
“No demand, so the company failed.”
The real issue is:
technology demand and social acceptance are diverging.
Why is resistance growing?
Flock’s biggest controversies relate to:
data volume,
search privileges,
how data is shared between agencies,
whether police should track vehicles long-term without a warrant,
and whether internal misuse occurs.
In September, SasaDaily reported multiple Indianapolis police officers accused of using Flock’s system for personal reasons instead of cases.
Now the bigger question is:
After repeated misuse, will markets reject the entire system?
Some places have already started limiting or shutting it down
Reuters reported that Florida has restricted Flock camera use on state highways due to privacy risks.
Virginia faces lawsuits about warrantless surveillance.
Recently, several US local governments began cancelling,
pausing,
or reviewing their contracts with Flock.
This means that as AI products scale in public spaces,
accuracy alone is no longer the only procurement criterion.
They must also address:
whether society accepts their constant presence.
Polls show no overwhelming public support
A Reuters/Ipsos poll showed:
38% support for community use of Flock cameras,
and 47% oppose.
This doesn’t prove all Americans reject AI license plate recognition,
but it does show it lacks broad consensus as a public technology.
For AI companies,
this creates a new commercial risk:
even effective products may face growth limits due to insufficient trust.
What should everyday people do?
No action is needed just because of Flock’s reported layoffs.
What matters more is that when encountering systems like:
AI cameras,
facial recognition,
license plate readers,
or workplace monitoring,
don’t just ask,
“Does it work?”
Also ask:
How long is data stored?
Who can access it?
Are there audit logs for queries?
What authorization is needed?
Can misuse be traced?
Can data be shared between agencies?
Because what really decides if such AI lasts long-term
is not how many vehicles it finds,
but:
whether people trust it won’t be abused.
3. Nvidia-backed Firmus withdraws $5 billion IPO: AI infrastructure is being asked “how will this pay off?”
The third case comes from Australia.
An AI data center company:
Firmus
was preparing a large IPO worth about
$5 billion.
If successful,
it would be among Australia’s largest-ever IPOs.
Firmus has heavyweight backers including:
Nvidia
and other major institutions and tech investors.
But on October 9,
Firmus officially withdrew its listing plan.
What’s Firmus’s explanation?
Firmus stated that due to
recent market volatility
and
prevailing market conditions,
current conditions do not reflect the company’s
business strength
and
long-term growth outlook.
Therefore, it has pulled the IPO and
will seek funding in the private market instead,
while considering other public or private options.
This is the company’s official position.
But what worries the market?
Reuters reported that investors have concerns about:
valuation,
execution risks,
operating history,
future capital needs,
and
when the massive AI infrastructure buildout will actually complete.
Initial IPO valuation was about
$30.6 billion,
roughly three times the company’s valuation just two months earlier.
For a company that must invest heavily in
GPUs,
power,
and data center construction,
the market is starting to demand
more than just
“AI demand will explode”
as the sole valuation rationale.
Firmus has big long-term goals, but actual capacity remains limited
Reuters Breakingviews noted that while Firmus’s long-term targets are huge,
its currently operational data center capacity
is only a small fraction of planned scale.
This reflects a reality many large AI infrastructure projects face.
In recent years, the market’s main question was:
“Is there enough compute power?”
Hence the rush to
buy GPUs,
borrow money,
and build data centers.
Now a new question arises:
“Who will pay for all this?”
And
“Can these assets generate enough revenue to pay back investors?”
This doesn’t mean the AI bubble has burst
That boundary must remain clear.
A company canceling an IPO
does not prove
the entire AI industry is collapsing.
Firmus itself has not halted operations.
The company says it will continue to
seek private capital,
and explore other market options.
Demand for AI compute power hasn’t vanished due to an IPO withdrawal.
So today’s key takeaway isn’t:
“The AI bubble has burst.”
But rather:
capital markets are becoming more discerning.
Why does this matter to the AI industry?
Because AI infrastructure costs
are now so large that
available cash from tech companies alone
may not suffice.
Many projects rely on a mix of
debt,
private equity,
project finance,
and public markets.
As investors demand
lower valuations,
more proof of cash flow,
clearer client contracts,
and more reasonable debt levels,
the pace of AI infrastructure expansion
could slow.
This may ultimately affect
GPU availability,
cloud pricing,
inference costs,
and even
the pricing of AI services themselves.
What should ordinary users do?
Nothing for now.
But if you operate an AI company or evaluate AI investments,
this development is a wake-up call to ask:
not just
“How big is the market?”
But also
how much capacity exists,
which clients have signed contracts,
which revenues are still planned rather than realized,
how much more debt/equity will be needed,
and if higher capital costs will threaten the business model.
AI is still a business.
Even massive demand doesn’t justify building infrastructure at any price.
The common turning point in today’s three stories
ARTEX shows us:
Having the capability doesn’t mean you can control who uses it.
Flock shows us:
A product’s effectiveness doesn’t guarantee social acceptance.
Firmus shows us:
Big demand doesn’t mean investors will pay any price.
Together, these stories show:
AI has entered its second stage.
The first stage asked:
“Can we build it?”
The second stage asks:
“Can we do it safely, gain acceptance, and sustain it economically?”
The AI companies and products that endure won’t be those with
highest-performing models,
biggest GPU counts,
or largest scale,
but those that also earn
control,
trust,
and
sustainable economics.
Recommended Reading
When Chip Sellers Start Guaranteeing to Buyers, Is the AI Boom Forming a Dangerous Cycle?
