Latest / AI Newspaper / China-Aligned Hackers Used AI Agents to Break Into Cloud Accounts in Under Three Hours, Microsoft Says

Transcript

China-Aligned Hackers Used AI Agents to Break Into Cloud Accounts in Under Three Hours, Microsoft Says · AI Newspaper

• 0:00
  Welcome to the AI Newspaper, your daily update
• 0:02
  on everything happening with AI. Visit AINewspaper
• 0:05
  .org to see today's stories and sign up for email
• 0:07
  alerts. So, imagine you lock up your office building
• 0:11
  for the night. You know a burglar might try to
• 0:14
  pick the front door, but, you know, you assume
• 0:16
  it's going to take them time. Time to test the
• 0:19
  locks, time to bypass the alarm, maybe time for
• 0:22
  a security guard to hear something and actually
• 0:24
  respond. Right, you're relying on that physical
• 0:26
  friction to give you a window. Exactly. But what
• 0:29
  if the burglar isn't a person at all? I mean,
• 0:32
  what if it's a machine that can test like 10
• 0:34
  ,000 different keys in a single second? And the
• 0:37
  moment it realizes the front door is just too
• 0:39
  secure, it instantly clones the digital badge
• 0:42
  of the building's health inspector and just walks
• 0:44
  right in. Yeah, right past the guards. Right
• 0:46
  past them. Today, we are looking at a genuinely
• 0:49
  terrifying new reality in cybersecurity. We are
• 0:52
  talking about an attack where the timeline from
• 0:55
  an initial target being selected to a complete
• 0:58
  Highly privileged network takeover shrank from
• 1:01
  months down to under three hours. It is. I mean,
• 1:06
  it's the absolute definition of a paradigm shift.
• 1:09
  We have spent the last few decades, basically
• 1:12
  our entire industry, building security operations
• 1:14
  centers around human speed attacks. Right. Because
• 1:18
  humans have limits. Exactly. Human hackers have
• 1:20
  to sleep. They take weekends off. You know, they
• 1:22
  have to manually process the data they steal
• 1:24
  to figure out what to do next. When they hit
• 1:27
  a wall, they have to literally stop and think
• 1:29
  about their next move. But what we are observing
• 1:32
  now with this specific incident is machine speed
• 1:36
  exploitation. And I want to really dig into the
• 1:40
  mechanics of how this happened, because it's
• 1:42
  not just about, you know, a computer being fast.
• 1:44
  It's about an off -the -shelf AI agent framework
• 1:46
  being hooked up to trusted third -party applications.
• 1:50
  So to really understand the gravity of this for
• 1:52
  the listener, we need to start with the threat
• 1:54
  group that pulled this off and the specific weapon
• 1:57
  they chose. Yeah. So the group in question is
• 2:00
  tracked by security researchers as Storm 0940.
• 2:04
  This is a China state aligned threat actor, and
• 2:08
  they've been active since at least 2021. Okay.
• 2:11
  Historically, they have targeted a really broad
• 2:14
  range of industries for data theft. We're talking
• 2:16
  aerospace, defense, energy, finance, government,
• 2:19
  you name it. But in the past, their methods were
• 2:23
  fairly traditional. What does traditional mean
• 2:25
  for a group like this? Well, they would do password
• 2:27
  spraying, which is basically guessing common
• 2:30
  passwords across hundreds of accounts. Or they
• 2:33
  would look for unpatched internet -facing applications
• 2:35
  to sort of sneak through. So traditional human
• 2:38
  -led hacking. Manual labor, essentially. But
• 2:41
  then in September of 2026, their methodology
• 2:44
  gets this massive automated upgrade. They start
• 2:48
  utilizing a tool called Artex. Tell me about
• 2:51
  this tool, because from what I understand, it
• 2:53
  wasn't actually built by hackers at all. No,
• 2:55
  and that is one of the most compelling or maybe
• 2:57
  frustrating parts of this entire situation. Artex
• 3:00
  was actually launched in July 2026 as an open
• 3:03
  source agentic AI penetration testing framework.
• 3:07
  Wow. Yeah, it was built for legitimate security
• 3:09
  teams. Let me just break down what agentic AI
• 3:12
  means in this context, because it's crucial.
• 3:14
  Yeah, please do. This isn't just a chat bot where
• 3:16
  you type a prompt and get an answer. It is a
• 3:19
  multi -agent system. So you have different AI
• 3:22
  models working together autonomously. One agent
• 3:25
  is tasked with reconnaissance. Another is tasked
• 3:27
  with attempting logins. Another analyzes the
• 3:30
  failure messages to adjust the strategy. So they
• 3:33
  are all talking to each other. Right. Yeah. It
• 3:36
  essentially mimics an entire team of human hackers
• 3:38
  communicating and adjusting their tactics in
• 3:41
  real time with almost zero human interaction
• 3:43
  required. And the architecture of this thing
• 3:46
  was quite sophisticated, right? I saw in the
• 3:48
  reporting it was written in Go with a Next .js
• 3:51
  front end. Now, I know a little bit about software
• 3:53
  development, but why does that specific tech
• 3:56
  stack matter to a cybersecurity defender? It
• 3:59
  matters because it tells you this wasn't just
• 4:00
  a quick script some kid hacked together in a
• 4:03
  basement. Go is a back -end programming language
• 4:06
  that is known for being incredibly fast and handling
• 4:09
  massive amounts of simultaneous connections.
• 4:11
  It's what we call concurrency. Okay, so we can
• 4:14
  do a lot at once. Exactly. When you want to launch...
• 4:17
  thousands of network requests a second without
• 4:19
  crashing your own computer you use something
• 4:22
  like go and the next js front end simply means
• 4:24
  they built a sleek highly professional web dashboard
• 4:27
  for it a user interface yeah a human operator
• 4:30
  could just sit back look at a clean interface
• 4:33
  point this ai swarm at a target and just click
• 4:36
  go which is exactly what happened i mean it was
• 4:38
  intended for good guys to test their own defenses
• 4:40
  but by october The creators actually had to pull
• 4:43
  it offline and make it closed source because
• 4:45
  it was already being weaponized. It was hitting
• 4:47
  South Korean financial firms and leaking over
• 4:50
  100 ,000 records. So Storm 0940 gets their hands
• 4:54
  on this, and they point it at Microsoft 365 environments.
• 4:57
  Right. Let's look at the sheer volume of this
• 5:00
  initial assault. This is between October 10 and
• 5:02
  October 14. The telemetry from those four days
• 5:05
  is just staggering. The AI framework launched
• 5:08
  a massive cross -tenant password spray, and they
• 5:11
  targeted 327 accounts across eight different
• 5:14
  organizations. In total, the security sensors
• 5:17
  captured over 151 ,000 failed logon events. Let
• 5:21
  me stop you right there, because the scale is
• 5:24
  hard to wrap my head around. The data shows they
• 5:27
  peaked at 2 ,900 failures per minute. In a single
• 5:30
  day, they had over 111 ,000 login failures. 48
• 5:34
  specific accounts received more than 1 ,000 login
• 5:37
  failures each. But here's where I'm genuinely
• 5:41
  confused and where I kind of have to challenge
• 5:43
  the idea that this AI was some unstoppable force.
• 5:46
  If I type my password wrong three times into
• 5:49
  my banking app, I get locked out. Yep, rate limiting.
• 5:52
  Right. And the data here explicitly shows nearly
• 5:54
  98 % lockouts. There was no successful authentication
• 5:57
  from this brute force, no tokens issued, no directory
• 6:01
  changes. The front door actually held up perfectly.
• 6:04
  So why are we calling this AI tool a game changer
• 6:07
  if it basically completely failed its primary
• 6:11
  mission? And that is the exact trap human defenders
• 6:13
  fall into when they analyze machine speed attacks.
• 6:17
  For a human hacker, 151 ,000 failures is a total
• 6:20
  disaster. It means you were noisy, you triggered
• 6:23
  every alarm, and your entire operation is blown.
• 6:26
  You'd pack up and run. Exactly. But for an agentic
• 6:29
  AI system... Failure is just data. The AI wasn't
• 6:33
  frustrated by the lockouts. It was learning.
• 6:35
  It was finding the exact threshold of the target's
• 6:38
  rate -limiting defenses. It blended broad, low
• 6:41
  -rate account guessing with these massive, high
• 6:44
  -volume bursts to basically map out the boundaries
• 6:47
  of the security system. Oh, wow. So it was testing
• 6:50
  the fences. Yes, testing the fences. And when
• 6:52
  it realized the front door was truly impenetrable,
• 6:54
  it didn't just give up and go home. It instantly
• 6:57
  pivoted to find a side door. And that pivot happened
• 7:00
  literally immediately. The very next day, October
• 7:03
  15th, using the exact same infrastructure, an
• 7:06
  IP address ending in 64 .99, the attack suddenly
• 7:10
  shifts completely away from password spraying.
• 7:12
  The AI agents realized that, you know, breaking
• 7:15
  the lock wasn't working. So they decided to just
• 7:17
  steal a master key instead. They targeted and
• 7:20
  successfully compromised a multi -tenant application
• 7:22
  called ConnectWise Cloud Backup. Okay. ConnectWise?
• 7:25
  Yeah. ConnectWise is what we call a managed service
• 7:28
  provider or MSP. They handle IT operations like
• 7:32
  data backups for thousands of small and medium
• 7:34
  -sized enterprises. Let's expand on this concept
• 7:37
  of multi -tenant applications because this is
• 7:40
  where the really crucial editorial lens of what
• 7:42
  we're looking at today comes into focus. We trust
• 7:45
  these third -party applications to hold our data,
• 7:47
  which means we give them incredibly high -level
• 7:50
  access to our systems. Precisely. A multi -tenant
• 7:53
  app means a single instance of the software serves
• 7:56
  multiple different customer organizations. Those
• 7:58
  are the tenants. Right. When a company installs
• 8:01
  a cloud backup solution, they have to grant that
• 8:03
  application permission to read. almost every
• 8:05
  file and directory in their entire network. Otherwise,
• 8:09
  you know, the backup just wouldn't work. The
• 8:11
  application is given an identity within the system,
• 8:14
  essentially a robot account that is highly, highly
• 8:17
  trusted. It brings me back to my earlier analogy,
• 8:20
  but let's take it even further. This isn't just
• 8:22
  sneaking past the bouncer at the front door.
• 8:25
  This is stealing the city health inspector's
• 8:27
  badge. I like that analogy. Because, I mean,
• 8:30
  every single restaurant in the city is legally
• 8:32
  obligated to let the health inspector straight
• 8:34
  into the kitchen, past all the locks, no questions
• 8:37
  asked, because they need the inspector to do
• 8:40
  their job so the restaurant can function. That
• 8:43
  is exactly what a cloud backup app is. It is
• 8:46
  the ultimate VIP entrance. Once Storm 0940 compromised
• 8:50
  those ConnectWise cloud application identities,
• 8:53
  the Artex AI agents completely abandoned the
• 8:56
  password guessing game. They didn't need to guess
• 8:58
  passwords anymore. They used the stolen app credentials
• 9:01
  to authenticate directly into the customer environments.
• 9:04
  And because the AI was driving the bus, the scale
• 9:07
  of this pivot was just as massive as the password
• 9:09
  spray. The agents attempted sign -ins to more
• 9:12
  than 2 ,250 Microsoft 365 tenants. And they maintained
• 9:17
  that exact same blistering pace, 2 ,900 attempts
• 9:20
  per minute. But this time, they weren't failing.
• 9:24
  No. They were walking right in. And this illustrates
• 9:26
  a truly chilling new paradigm for the industry.
• 9:29
  When an off the shelf AI agent framework lets
• 9:32
  a state aligned hacking group go from first targeting
• 9:35
  to privileged cloud access in under three hours
• 9:37
  across hundreds of organizations, defense stops
• 9:40
  being about spotting the alert and becomes about
• 9:44
  acting on it at machine speed. Right. And every
• 9:47
  trusted third party app with broad permissions
• 9:49
  basically becomes a door that opens just as fast.
• 9:53
  Organizations spend millions on securing their
• 9:55
  human employees, right? Multi -factor authentication,
• 9:58
  complex passwords, security training. Don't click
• 10:00
  the phishing link. Exactly. But all of that is
• 10:03
  bypassed when an attacker hijacks the silent,
• 10:06
  automated trust we give to these vendor applications.
• 10:09
  So how does an organization even begin to defend
• 10:12
  itself when the call is coming from inside the
• 10:15
  house? The security system sees the backup app
• 10:18
  logging in and says, oh, well, let's just...
• 10:20
  the backup app doing its daily job. There's no
• 10:22
  red flag. It is incredibly difficult to detect
• 10:25
  initially because, as you said, these applications
• 10:27
  are designed to operate without interactive human
• 10:30
  sign -ins. The AI agents were taking advantage
• 10:33
  of trust that had already been established within
• 10:35
  the Microsoft tenant. The time between the initial
• 10:38
  compromise of ConnectWise and the widespread
• 10:40
  exploitation of its downstream customers was
• 10:43
  virtually non -existent because the AI executed
• 10:45
  the pivot entirely automatically. Which brings
• 10:49
  us to the truly terrifying part of this whole
• 10:51
  thing, which is what these AI agents actually
• 10:53
  did once they were inside those Microsoft 365
• 10:55
  tenants. I mentioned at the top of the show that
• 10:58
  the timeline from targeting to full privileged
• 11:01
  access was under three hours. Under three hours,
• 11:05
  yeah. Think about what three hours is in a normal
• 11:07
  corporate IT department. That's like barely the
• 11:10
  morning status meeting in a coffee break. By
• 11:13
  the time the IT team is sitting down at their
• 11:15
  desks, the entire network is already compromised.
• 11:18
  And all of it happened without a human hacker
• 11:21
  typing a single command. Once inside, the AI
• 11:24
  agents began conducting autonomous reconnaissance
• 11:26
  using the Microsoft Graph API. Now, I hear a
• 11:29
  Graph API thrown around a lot in these security
• 11:32
  reports. Can you translate what that actually
• 11:35
  is in plain English? How was the AI using it?
• 11:37
  Sure. Think of the Microsoft Graph API as the
• 11:40
  central nervous system of a company's entire
• 11:43
  digital workspace. It is the programming interface
• 11:46
  that connects emails, chat messages, file directories,
• 11:49
  user profiles, organizational charts, everything.
• 11:52
  If you want to know who the CEO is, what files
• 11:54
  they accessed yesterday and who they emailed,
• 11:56
  you query the Graph API. The AI agents were using
• 12:00
  this to instantly map out the entire corporate
• 12:03
  structure, figuring out exactly what resources
• 12:05
  they had just gained access to. They are standing
• 12:08
  in the lobby with the building blueprints, instantly
• 12:11
  knowing where every sensitive room is. Yes. And
• 12:14
  the data shows they immediately headed for the
• 12:15
  most sensitive rooms, which are the Azure Key
• 12:18
  Vaults. Azure Key Vaults are essentially the
• 12:20
  master digital safes of a cloud environment.
• 12:23
  It is where a company stores its most critical
• 12:25
  secrets. We're talking encryption keys, API tokens,
• 12:29
  administrative passwords for other databases.
• 12:32
  The crown jewels. The absolute crown jewels.
• 12:35
  The telemetry showed bulk secret reading across
• 12:38
  multiple vaults by a single identity. The AI
• 12:41
  was performing key enumeration, which is basically
• 12:43
  asking the vault to list every single secret
• 12:45
  it holds, and then key unwrapping operations.
• 12:48
  And unwrapping means? Unwrapping is the process
• 12:51
  of decrypting those secrets so they can actually
• 12:53
  be used by the attacker. Okay, so the AI has
• 12:55
  mapped the building, it has opened the master
• 12:57
  safe, and has copied all the passwords. But then
• 13:00
  it goes a step further into what the security
• 13:02
  reports call... persistence. And I want to unpack
• 13:06
  that term because it sounds clinical, but the
• 13:09
  reality of what the AI was doing is, well, it's
• 13:11
  insidious. It wasn't just stealing data. It was
• 13:14
  fundamentally altering the locks from the inside.
• 13:17
  Persistence is the holy grail for an attacker.
• 13:21
  It means ensuring that even if you are discovered
• 13:23
  and kicked out, you don't lose access. The AI
• 13:27
  agents began modifying application credentials
• 13:29
  within the targeted victim's environments. They
• 13:32
  were creating what we call new service principles.
• 13:35
  What is a service principle, exactly? A service
• 13:38
  principle is simply the local identity or the
• 13:41
  account that an application uses to run within
• 13:43
  your specific network. Like we said earlier,
• 13:46
  it's a robot user. Right. The AI was creating
• 13:49
  new robot users and modifying existing ones.
• 13:52
  Most notably, they were exploiting OAuth app
• 13:55
  permissions. OAuth, like when you log into a
• 13:57
  site with your Google account? Exactly. It's
• 14:00
  the same underlying technology that lets you
• 14:02
  log in with Google on a random website. But for
• 14:06
  enterprise apps, it grants deep systemic permissions.
• 14:09
  The AI manipulated these settings so that the
• 14:12
  compromised application self -assigned highly
• 14:14
  privileged roles to itself. Roles like what?
• 14:18
  Like global administrator. Wait, global administrator.
• 14:22
  So it gave itself the highest possible rank in
• 14:24
  the entire IT department. It essentially crowned
• 14:27
  itself king of the network. Yes. It granted itself
• 14:31
  new access tokens that are entirely separate
• 14:33
  from the original ConnectWise vulnerability.
• 14:36
  Which leads to a really terrifying realization.
• 14:38
  Let's say a company's IT admin reads the news,
• 14:41
  sees that ConnectWise is compromised, and immediately
• 14:44
  disables the backup application. They probably
• 14:46
  think they are safe. But they aren't, are they?
• 14:50
  Exactly how long can these new AI -created backdoors
• 14:54
  lay dormant in the system? Indefinitely. Wow.
• 14:57
  And that is the true danger of automated persistence.
• 15:00
  These credential modifications remain active
• 15:02
  as a backdoor long after the original compromise
• 15:05
  application is remediated. The AI has cut its
• 15:08
  own master keys. Going back to your analogy,
• 15:11
  disabling the health inspector's badge doesn't
• 15:13
  matter if the health inspector already came inside,
• 15:16
  hired three new employees, and gave them their
• 15:18
  own keys to the backdoor. It is important to
• 15:20
  note, though, based on the reporting, that while
• 15:22
  this attack targeted a massive cross -section
• 15:25
  of industries globally, aerospace, defense, government.
• 15:28
  The mere presence of the compromised application
• 15:30
  doesn't automatically mean data was successfully
• 15:32
  stolen from every victim. The security alerts
• 15:35
  emphasize that successful authentication doesn't
• 15:38
  always equal data exfiltration. That is a crucial
• 15:42
  distinction. Just because they broke into the
• 15:44
  safe doesn't mean they had time to carry all
• 15:46
  the gold out before someone noticed. Right. But
• 15:49
  the potential exposure is catastrophic. When
• 15:51
  an attacker has a global administrator role,
• 15:53
  they have the keys to the entire kingdom. And
• 15:56
  because the AI is identifying and exploiting
• 15:58
  that access at 2 ,900 attempts a minute, human
• 16:02
  defenders are completely outmatched. Which forces
• 16:04
  us to look at the defensive realities of this
• 16:06
  new landscape. If an attack goes from initial
• 16:10
  targeting to full privileged access across hundreds
• 16:13
  of organizations in under three hours, the traditional
• 16:16
  security operations center, the SOC, is essentially
• 16:20
  obsolete, isn't it? It fundamentally breaks the
• 16:22
  traditional model. A typical SOC. operates on
• 16:25
  a triage system. An alert pops up on a dashboard.
• 16:28
  A human analyst looks at it, determines if it's
• 16:31
  a false positive, and then escalates it. But
• 16:33
  you can't triage that fast. Exactly. Human analysts
• 16:36
  cannot triage 2 ,900 alerts a minute. Defense
• 16:40
  stops being about spotting an alert on a dashboard
• 16:42
  and becomes about acting on that alert at machine
• 16:45
  speed. By the time a human analyst even opens
• 16:47
  the first support ticket, the AI agent has already
• 16:50
  mapped the directory via the graph API, self
• 16:52
  -assigned global admin rights, and extracted
• 16:54
  the Azure key vaults. So the entire operational
• 16:57
  reality for defenders has to change. And this
• 17:01
  leads to a really counterintuitive piece of advice
• 17:04
  in the incident response guidelines. The immediate.
• 17:07
  gut reaction of any IT professional when they
• 17:10
  hear a specific app is compromised is to just
• 17:13
  go into the system and delete it, you know, kill
• 17:15
  it with fire. Right. But the security guidance
• 17:18
  explicitly tells defenders not to do that as
• 17:21
  a first response. Why? Because deleting the enterprise
• 17:25
  application is the equivalent of destroying a
• 17:27
  crime scene before the detectives even arrive.
• 17:29
  If you delete the app, you wipe out vital configuration
• 17:32
  data and telemetry. You have to preserve the
• 17:35
  data in your SIEM, your security information
• 17:38
  and event management system. That is your centralized
• 17:41
  security dashboard, essentially the flight data
• 17:43
  recorder for your network. So if you delete the
• 17:46
  app, you smash the flight recorder. Exactly.
• 17:48
  You lose the Entra sign -in logs, Entra being
• 17:51
  Microsoft's identity and access management system.
• 17:53
  You lose the audit records, the graph API telemetry.
• 17:57
  You lose the breadcrumbs that tell you which
• 17:59
  new keys the AI cut. I see. The operational reality
• 18:03
  dictates that connections must be disabled to
• 18:05
  stop the bleeding, but the configuration must
• 18:08
  be preserved for forensics. You cannot delay
• 18:10
  urgent containment, but you can't destroy your
• 18:13
  own visibility either. And defenders have to
• 18:15
  completely shift their focus when they review
• 18:17
  these logs, right? It's not just about looking
• 18:19
  at which human employees are logging in from
• 18:21
  weird locations anymore. Not at all. Defenders
• 18:24
  have to review the service principal sign -ins,
• 18:26
  those automated robot accounts. They have to
• 18:28
  separate the successful sign -ins from the failed
• 18:30
  ones, looking for unfamiliar IP addresses and
• 18:33
  activity that happens outside the expected vendor's
• 18:36
  infrastructure. Looking for anomalies. Yeah,
• 18:38
  they need to heavily scrutinize OAuth app permissions,
• 18:41
  reviewing audit records for any new credentials
• 18:44
  added to service principals and any new privileged
• 18:47
  role assignments that happened right around the
• 18:49
  time of the compromise. It sounds like a massive...
• 18:52
  Needle in a haystack forensic undertaking. Which
• 18:55
  leads me to a really philosophical but urgently
• 18:58
  operational question. I'm thinking about this
• 19:01
  timeline again. If a smoke alarm goes off, but
• 19:05
  the house burns to the ground in exactly three
• 19:07
  seconds, what good is the alarm? Are we at the
• 19:10
  point where organizations have to hand the keys
• 19:12
  of their cybersecurity defense entirely over
• 19:14
  to AI agents just to keep up with the attacking
• 19:18
  AI agents? I believe that is the defining question
• 19:21
  for the next era of the cybersecurity industry.
• 19:24
  Right now, attackers are utilizing AI to exploit
• 19:27
  the trust built into our supply chains faster
• 19:29
  than humans can possibly revoke it. When every
• 19:33
  trusted third -party app with broad permissions
• 19:35
  becomes a door that opens at machine speed, human
• 19:38
  oversight just becomes a fatal bottleneck. It's
• 19:40
  too slow. Way too slow. We are rapidly moving
• 19:43
  toward a reality where AI -driven defense platforms
• 19:46
  are the only systems capable of detecting a 2
• 19:49
  ,900 attempt per minute anomaly and automatically
• 19:52
  isolating the compromised tenant before the third
• 19:55
  minute even begins. It really makes you stop
• 19:57
  and think about the nature of the tools we are
• 19:59
  building in this industry. I mean, the framework
• 20:02
  at the very center of all of this, Artex, was
• 20:05
  originally built and released as an open source
• 20:07
  project to help companies test and improve their
• 20:10
  own security. It was intended to be a shield.
• 20:13
  It was built to help good guys find their own
• 20:15
  vulnerabilities before the bad guys did. But
• 20:17
  if the open source tools we build to defend ourselves
• 20:20
  are instantly co -opted and weaponized by state
• 20:24
  -aligned groups to attack us at machine speed,
• 20:27
  are we just trapped in an endless loop? Like,
• 20:30
  are we entering an arms race where AI forces
• 20:32
  both sides to move faster and faster and faster
• 20:35
  until human beings are entirely out of the loop
• 20:38
  on both the attack and the defense? It is a chilling
• 20:41
  thought to leave you with, but it is the unavoidable
• 20:44
  reality of the landscape we now operate in. The
• 20:46
  speed of conflict has permanently changed. Thanks
• 20:49
  for listening to the AI Newspaper. Visit AINewspaper
• 20:52
  .org to catch up on the latest updates and sign
• 20:55
  up for free email alerts.