Last Week in Tech - F5 and Check Point Under Fire, a $387 Million Spoofed Approval, and AI Agents Go Carding
23 minAudio
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Last Week in Tech for Monday 28 September 2026, covering the week from 21 September. F5 BIG-IP APM zero-day (22 Sep). CVE-2026-94127, heap overflow enabling unauthenticated RCE. F5 scores it CVSS v3.1 9.8 and CVSS v4.0 9.3. Exploitable only where APM acts as an OAuth authorisation server with an access policy and OAuth profile on the same virtual server. Added to CISA KEV the same day, federal deadline 25 Sep. Hotfixes for 17.1.x, 17.5.x and 21.1.0, plus an iRule workaround via F5 Support. Shadowserver sees 14,700+ IPs with a BIG-IP APM fingerprint; patch status unknown. Check Point firewalls and management (22 Sep). CVE-2026-85102, pre-auth RCE in Security Gateway and Spark VPN certificate handling, patched 9 Sep, with exploitation attempts against Spark customers since 12 Sep. CVE-2026-93616, pre-auth path traversal in Security Management, a zero-day with targeted attacks seen on 23 July and a fix released 22 Sep; LivePatch does not fix it. Both rated critical by Check Point. Also in the same KEV batch: Arista VeloCloud Orchestrator CVE-2026-93952, actively exploited, rated critical; fixed in 5.2.3.16 and 6.4.2.8. Bitget hack (24 Sep). First put at US$351.6M, revised by CEO Gracy Chen on 25 Sep to about US$387.5M after uncounted transfers from the same incident were traced. Attackers compromised a wallet backend, spoofed transaction data and triggered Bitget's own authorisation process; private keys were not taken. Hot and warm wallets hit; cold wallets secure. Losses covered by Bitget's User Protection Fund; some funds frozen; 5% recovery bounty offered. Bitget says the flaw is fixed and is reopening withdrawals in phases from 28 Sep. Bitget suspects North Korea - its own assessment, not confirmed by law enforcement. AI agents run a carding campaign (single-source). Gambit Security describes an ongoing campaign dating to July in which one operator used three open-source agent tools. Between 10 and 15 Sep alone it counted 105 attacks and at least 27 unnamed companies compromised; 600,000+ card records were taken from two victims, skimmers confirmed on 19 named victims, and 100+ further infected sites linked. Mean cost about US$25 per completed scan; campaign estimated at US$12,000-18,000. Akamai-Anthropic deal (24 Sep). US$11.6bn over seven years for CPU workloads on Akamai Cloud, expandable by up to US$9bn. Anthropic receives a warrant for up to about 5% of Akamai. Akamai expects about US$5.5bn in related capex, partly to pre-buy memory. ShinyHunters' FBI claim. The FBI confirmed only that it is aware of a claimed compromise of FBIJobs.gov and is investigating. The PeopleSoft zero-day, data volume and systems named are unverified claims; no new CVE. PeopleSoft users should confirm the June fix for CVE-2026-35273 is applied. Coming up: Wednesday, the OpenAI agent inside a Services Australia portal. Friday, one stolen app key and hundreds of online stores. Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: F5 advisory K000162605; CISA KEV; BleepingComputer; The Register; Rapid7; Check Point; Arista SA-0183; SecurityWeek; CoinDesk; CNBC; Gambit Security; Akamai newsroom; TechCrunch; Cybersecurity Dive.
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Welcome to today's deep dive. Imagine for a second spending millions of dollars on state -of -the -art digital security, right? You fortify your network, you hire the best teams, only to realize that the very firewalls you bought to protect you are like the exact doors hackers are using to break in. Yeah, it completely flips the entire concept of security on its head. It really does. And we know you are here because you want a clear, you know, hype -free synthesis of exactly this
kind of paradigm shift. You're looking for the actual signal through the noise of the weekly news cycle. Right. And it is a profound shift. Honestly, it's the definitive theme we pulled from our briefing covering the most critical tech and cybersecurity events from the past week, ending September 28th, 2026. Which is just a packed week. Oh, yeah, absolutely. And the core takeaway is that the foundational technologies and systems we just implicitly trust to keep us safe, well, they're failing at a structural level. Yeah. And our mission today is to explore
exactly how this illusion of control is being shattered for you. We're going from perimeter defenses becoming the actual attack surface to a massive crypto heist where the security system functioned perfectly but still authorized the theft. And then looking at how the sheer economics of cybercrime are just collapsing due to AI. It's a lot of ground to cover. It is. So let's start right at the perimeter wall. We historically view firewalls and VPN gateways as our digital armor. But this week, the major story is about
vulnerabilities in F5, Big IP, and checkpoint systems. It's like the armor essentially grew spikes on the inside. Right. And the speed of the response from federal authorities here really tells you everything you need to know about the severity. So on September 22nd, F5 published an advisory for a flaw in their big IP access policy manager. And this isn't just a small tool, right? No, not at all. This is a component massive organizations use. Think global banks, governments, Fortune 500s. They all use it to manage remote access. And the U .S. Cybersecurity and Infrastructure
Security Agency, or CISA, they added it to their known exploited vulnerabilities catalog that exact same day. Wow, the exact same day. Yeah, giving federal agencies a mere three days to secure it. Okay, let's unpack this. A three -day turnaround for federal agencies is a massive scramble. I mean, the vulnerability is cataloged as CVE -2026 -94127, and it carries a 9 .8 severity score. out of 10. Nearly perfect. Right. Almost as bad as it gets. The briefing describes it
as an unauthenticated remote code execution caused by a heap buffer overflow. And I know we've touched on buffer overflows before, but remind us of the mechanics behind why this allows someone to just walk in without a password. Well, to understand the mechanics of a heap buffer overflow, you have to think of a computer's act of memory, the heap, as a series of meticulously organized, flexible storage bins. Okay. Bins. Got it. Right. So a buffer overflow happens when an attacker intentionally shoves way too much data into one
specific bin. The software basically fails to check the size of the data, so the excess spills over into the adjacent bins. And what's in those adjacent bins usually? Well, that's the danger. If those adjacent bins happen to hold critical system instructions, the attacker's spillover data actually overwrites those instructions with their own malicious code. Oh, wow. Yeah. So the system then executes that malicious code thinking it's just part of its normal routine. So it's not even like a broken lock on a bank vault.
It's more like a bank teller's cash drawer. Like if a hacker shoves a massive stack of $1 bills into the $20 slot and they spill over into the $100 slot, suddenly the system's reality is warped. It thinks you deposited hundreds instead of ones. That is a highly accurate way to visualize it, actually. You've just overwritten the system's reality by cramming too much data into a space that wasn't built to contain it. Exactly. You are manipulating the system's memory to force it to execute commands it was never programmed to run. And what's fascinating here is that we
are witnessing the devices meant to face the hostile Internet and protect our internal networks becoming the primary attack surface themselves. Because they're the things exposed to the outside. Right. And we saw a similar crisis unfold on the very same day with Checkpoint. Yeah. Checkpoint released an advisory bluntly titled Action Required, dealing with two critical flaws being actively exploited. One involved how their security gateways and Spark firewalls handle VPN certificates. But the second one really caught my eye. The
zero day. Yes, it was a true zero day in their security management server. It was described as a pre -authentication path traversal vulnerability. Wait, let me make sure I've got path traversal right. Is that like if I ask a librarian for a specific book in the public reading room, but I use some weird coded language that tricks her into walking backwards through the staff doors and handing me the master keys to the building? You've captured the exact mechanism there. In a typical web application, a user might request a file. Like an image. And the application looks
in the public images folder. Makes sense. But with a path traversal vulnerability, the attacker inputs a sequence of characters, usually a series of dots and slashes. And that commands the server to move backwards out of that public directory and up into the restricted system directories. So they're literally just telling the computer to go back a folder? Yes, over and over. And suddenly, instead of serving a harmless image, the server is handing over password files or execution scripts. And importantly for checkpoint
users, fixing this requires a full system update. Oh, you can't just patch it. No, you can't just apply a quick live patch, which is their system for applying fixes without rebooting. It requires actual downtime. And just to show how widespread this perimeter failure really is, Arista's Valero Orchestrator also has a critical, actively exploited flaw right now regarding certificate authentication. Right, the orchestrator. Yeah, which manages their SD -WAN products. For anyone unfamiliar, that stands for Software Defined Wide Area Network.
It's essentially the intelligent traffic cop that routes a company's data across different internet connections. If that gets compromised, well, the attacker controls the map. They absolutely do. But to bring this back to practical reality for you, the listener, There is a crucial nuance with the F5 situation specifically. Okay, what's that? It is only exploitable if the access policy manager is configured as an OAuth authorization server. Okay, wait, let's define that quickly because that sounds like a massive piece of jargon that dictates whether an IT team needs to panic
today or not. Yeah, fair enough. OAuth is essentially the digital bouncer for your network. When you click log in with Google or log in with Microsoft on a third -party website, That's OAuth. You are giving the third party your password, right? The bouncer verifies your identity and hands the third party a temporary VLP wristband, a token that grants them access. Oh, I see. So if your F5 system is set up to be that bouncer generating those wristbands, you are vulnerable. Scanners show over 14 ,700 internet -facing IP
addresses with that specific big IP fingerprint right now. Over 14 ,000. So the strategic takeaway for anyone running F5 or Checkpoint Spark isn't just to casually patch during the next cycle. The consensus from the data is clear. You have to assume you've been probed. The most critical question for IT teams this morning is checking for that specific OAuth bouncer configuration. Precisely. Now, if the perimeter walls, the firewalls, and VPNs are compromised, attackers are inside the network. So how do we catch them? Right.
Usually we rely on human verification and internal controls, but our next piece of intelligence really demonstrates what happens when the information humans rely on to verify transactions is structurally compromised. Which brings us to the BitGet cryptocurrency exchange theft. On September 24th, BitGet detected unauthorized transfers that resulted in a staggering $387 .5 million being stolen. Just a massive amount of money. Huge. But the amount isn't even
the most alarming part. It's how they did it. The attackers didn't steal the private keys to the crypto wallets. Instead, they compromised a back -end system, spoofed the transaction data, and fed that fake data into BitGet's own authorization process. The methodology is the core issue here. The internal controls worked exactly as they were designed to work. Right. The system required an approval, and an approval was given. Exactly. The problem is that the attacker controlled the back end, preparing the request. See, here's
where it gets really interesting for me. I'm reading this and thinking, a human still had to approve this transfer. If a human clicks approve on sending hundreds of millions of dollars to a criminal, how is that not just a colossal human error? I mean, shouldn't they have noticed the destination address was wrong? It is entirely natural to look for a human scapegoat in that scenario. But it is not human error if the screen is actively lying to the human. Lying to them? Yeah, think about it. If the human is looking at a dashboard that says, transfer $50 ,000 to
internal wallet A for daily liquidity, they will click approve. But if the attacker has compromised the backend API, the middleman that carries data from the database to the screen, they can alter the reality. How exactly does that API interception work, technically, though? So the front end, which the human sees, asks the back end database for the transaction details. The attacker is sitting right in the middle, intercepting that request. Okay. The database says, yes, this is a routine $50 ,000 internal transfer. The attacker's
script grabs that data before it reaches the screen, rewrites it to still say routine $50 ,000 transfer, but changes the hidden destination code to the attacker's wallet. Oh, man. And they multiply the hidden amount. So the human sees a perfectly legitimate looking screen, clicks approve, and the system logs that a human verified the transaction. The human control is entirely bypassed because they were just rubber stamping a lie. So they didn't forge the signature on the bank check. They altered the printed numbers on the check after it was already signed, essentially.
Basically, yes. Just to wrap up the facts on BitGet, they claim their cold wallets are secure, their user protection fund covers the loss, and they've actually started reopening withdrawals. Their internal assessment suspects North Korea is behind it, based on VPN IP addresses, though we always have to note that's unverified by law enforcement at this stage. They've also frozen some funds and offered a 5 % recovery bounty. The larger implication here translates directly into your daily business workflows, far beyond cryptocurrency. How so? Well, consider your own
finance team. When they approve a massive payroll run or a vendor payment into your accounting software, where does the information on their screen actually come from? Could a compromised API alter the destination routing numbers before the human ever sees the final approval screen? That is a terrifying thought. We inherently trust internal dashboards. But if the foundation feeding them is compromised, your human verification is totally useless. And bypassing human controls like that requires intense sophistication. Doing
it at scale used to require an army of highly skilled, expensive hackers. But looking at our next source, the barrier to entry for that level of sophistication has completely vanished. It really has. The economics of cybercrime are just collapsing due to AI. We are pulling from a sobering report by Gambit Security regarding a Chinese -speaking operator running an AI card skimming campaign. And the operator didn't write hundreds of custom exploits. They used three open source AI agent tools to build an automated self -managing
attack pipeline. The technical breakdown of this is wild. They used one AI agent as the orchestrator, like the project manager. A second AI agent was the scout. deployed to scan the internet for vulnerabilities. And a third AI agent was the demolition expert executing the actual exploits. And the efficiency is staggering. Between September 10th and 15th, they executed 105 attacks. They compromised at least 27 companies, including a Fortune 500 hospitality brand and a major U .S. airline. Wow. Over 600 ,000 credit card records
were stolen from just two of those victims. I have to push back here, though. Aren't models like Anthropix Clod or OpenAI systems supposed to have strict safety guardrails precisely to stop them from writing malicious code? This raises an important question, and the source material provides a fascinating look into the attacker's methodology here. The attacker used Anthropix Clod Opus 4 .6 for the orchestration agent. Okay. The report specifically notes that newer models did refuse the attacker's request due to safety guardrails. So the attacker essentially went
model shopping. They utilized an older or jailbroken version of Claude to act as the brains, while using models like DeepSeek and GLM for the grunt work. Okay, let's unpack this orchestration part, because how does a chatbot actually orchestrate an attack? It's not just sitting there typing out code into a terminal, right? No, no, these aren't chatbot interfaces. These are API agents plugged directly into the execution environment. Claude acts as the manager. It is given a high -level goal like extract data from this target.
And then what does it do? It generates the command prompts, sends them to the deep -seek scanning agent, reads the scout's report, and then formulates a specific attack plan. If the GLM execution agent tries the attack and gets an error code back from the server, the cloud orchestrator reads the error, understands why the attack failed, writes a new script to bypass the error, and orders the execution agent to try again. Wait, so it just fixes its own mistakes on the fly? Yes, it is dynamically adjusting the attack in
real time without any human intervention. That is terrifying. And because they are using automated agents, the cost is basically a rounding error. The report estimates the mean cost per completed scan was about $25. Yep, $25. The entire campaign, which yielded hundreds of thousands of stolen credit cards, cost between $12 ,000 and $18 ,000 total. It shatters the economic model of cyber defense. A criminal can now buy the equivalent of a highly skilled 24 -7 hacking team for the
price of a used car. So what is the defense against a $25 AI mercenary then? The briefing points specifically to compliance standards, namely PCI DSS version 4. For anyone who doesn't handle e -commerce, that's the payment card industry data security standard. Historically, a lot of merchants viewed PCI compliance as just... you know, an annual paperwork headache. Right, but version 4 introduces strict requirements to actively monitor the scripts running on your payment pages. Previously, you might review your code every
few months. Sure. But with the $25 AI bot constantly probing your site, capable of dynamically rewriting its attack to bypass your defenses in seconds, automated monitoring is no longer just a paperwork exercise. It's your only line of defense fast enough to notice an AI injecting a malicious card skimmer into your checkout page before your customer's data is exported. The speed of AI attacks mandates high -speed automated defense. But running 10 to the pi of dynamic AI agents
a week isn't magic, right? It requires massive physical computing power. And that insatiable demand for processing power is quietly triggering a massive squeeze in the global hardware supply chain. which brings us to an incredible infrastructure shift. Yeah, we are looking at an $11 .6 billion seven -year agreement between Akamai and Anthropic, with room to expand by another $9 billion. Huge numbers. Anthropic gets CPU workloads on Akamai's cloud, and in return, Anthropic gets a warrant for up to roughly 5 % of Akamai's stock. And
in response to securing this deal, Akamai expects about $5 .5 billion in related capital spending. Crucially, they are pre -buying $1 .7 billion in physical memory this year alone to build out the capacity. Wait, Akamai, I know them as a content delivery network and a security vendor. Their main job is making sure websites load fast across the globe and blocking DDoS attacks, which is a distributed denial of service. Basically, when hackers flood a website with so much junk traffic, it crashes. Why is a network routing
and security company getting into the heavy -duty AI computing business? Because the sheer physics of AI demand it. Anthropic needs compute power closer to the edge of the network, and Akamai has data centers everywhere. We are seeing traditional security vendors morphing into massive compute businesses just to feed the AI beast. So what does this all mean? If a company whose main job is routing internet traffic is suddenly buying... billions in physical memory chips. It's like a local parcel delivery service buying up all the diesel fuel in the entire country because
they just partnered with a global freight train network. That analogy perfectly captures the market disruption. When a single company pre -buys $1 .7 billion worth of memory, it sends a massive signal. Prices are going up. Hardware is going to get scarce and it is going to get expensive. The AI boom is literally hoarding the world's computer memory. Which means the strategic takeaway for you listening is immediate. If your company has a server refresh scheduled or you need to buy a fleet of new laptops for your team in the next 12 months, you need to
call your vendors and lock in your pricing right now. Do not wait or you will be competing against these billion dollar AI infrastructure build outs. We are looking at a wild future of dynamic AI agents and hardware shortages. But, you know, the irony of our digital foundation is that some of the most critical data in the world. isn't being guarded by futuristic tech. It's sitting on legacy software, which is just as vulnerable. Right, which brings us to the extortion claims involving the FBI. The cybercriminal group known as Shiny Hunters claimed this week that they
breached the FBI. They say they used a new unpatched zero -day vulnerability in Oracle PeopleSoft to steal terabytes of data about FBI staff and job applicants. As your guides through this intelligence, it is vital we clearly delineate confirmed fact. from criminal claim here. Always. The factual baseline is this. The FBI confirmed they are investigating a claimed compromise. The FBIjobs .gov portal was taken offline, and the FBI stated
that the point of compromise, whether it was a third -party vendor or their own systems, is still undetermined. Exactly. The claims of a new Oracle Zero Day, the specific scale of the stolen data, and the actual systems involved are strictly the extortion group's unverified claims. Oracle has not commented, and there is no new CVE issued for this supposed Zero Day. Right. But I have to ask, if the Zero Day and the terabytes of data are completely unverified
claims from criminals trying to drum up extortion leverage, Why are we dedicating time to it on today's deep dive? Why give them the oxygen? If we connect this to the bigger picture, whether Shiny Hunter's specific claim is true or a complete fabrication, the threat vector they are pointing at is incredibly real. Oracle PeopleSoft is a ubiquitous legacy HR system. It is heavily relied upon by government agencies, massive corporations, and universities everywhere. Ah, I see. So while
everyone is worried about highly advanced AI orchestrators, the backdoor is a 15 -year -old HR portal processing job applications. It's the foundation cracking again. You might have a cutting -edge AI firewall, but your legacy administrative systems are your actual attack surface. Precisely. And the takeaway here is clear. If your university or government agency runs PeopleSoft, you cannot wait around to see if the zero -day rumors are true. What should they do? You must ensure that the confirmed June PeopleSoft fix, that is CVE
-2026 -35273, is applied immediately to remove the known vulnerabilities while we wait for the facts on the unknown ones. Okay, let's take a breath and recap the journey we've been on today. We set out to explore how the foundations of digital trust are shifting. We saw how the very firewalls meant to protect us, like F5 and Checkpoint, are now the primary targets due to deep memory flaws and path traversals. We saw how human approval processes mean nothing if a crypto platform's
backend API is spoofing the data right on the screen. We looked at how a $25 AI cyber mercenary operation can dynamically rewrite code to steal hundreds of thousands of credit cards, and how the AI boom is literally hoarding the world's computer memory. The illusion of control is dissolving across the board. We are relying on legacy software, vulnerable perimeters, and screens that can literally lie to us. Which leaves us with a final, rather provocative thought. Lay it on us. If an AI can
now orchestrate a highly complex cyber attack for the price of a takeout dinner, and if our human approval processes can be so easily bypassed by spoofed backend data, What happens when those two realities merge? Oh, man. What happens when an AI agent doesn't just spoof the financial data on your dashboard, but also spoofs the internal emails, the Slack messages, and generates the voice memos from your boss asking you to approve it? That is wild. How will we verify what is
actually real inside our own companies when the foundation of communication itself can be so cheaply and perfectly fabricated? That is a heavy, heavy question to mull over. If the digital foundation we stand on is shifting that rapidly, maybe it's time we stop admiring the view from the penthouse and start thoroughly inspecting the basement. Thank you so much for joining us on this deep dive. Keep questioning the systems you trust, keep learning, and we'll see you next time.
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Japanese home goods company Iris Ohyama (online store operator Iris Plaza) found a suspicious file in an external...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Australian online retailer Who Gives A Crap told customers on 7 November 2026 that their information was accessed...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Last week in tech, 7 to 14 November 2026. Microsoft's November Patch Tuesday (10 November) fixes well over a...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Chainalysis disclosed on 30 October 2026 that an external actor using "offensive agentic techniques" accessed...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! NSW Health reported unauthorised access to its Oral Health Fee For Service Scheme website on 29 October 2026;...