Field Notes
Evidence from the field
Notes from the work itself. How attackers really get in, what we find when we go looking, and the fixes that hold. No theory, no scare copy. Just the walk-through.
- Threat Intelligence
GTA 6 launches Thursday. The fake PC download is already waiting, and half your IT team will be playing.
Grand Theft Auto VI comes out on 19 November for PlayStation 5 and Xbox only. Every “GTA 6 PC download” is fake, and researchers have pulled infostealers, remote-access trojans and a wiper out of them. What they steal from a work laptop, why the launch weekend is thin on defenders, and seven things to do before Thursday.
8 MIN READ - Outage
Microsoft 365 went down for more than eleven hours. Your email security went down with it.
On 12 November Outlook, Teams, SharePoint and Copilot failed for users in 74 countries. Fewer people noticed that Defender, Purview and Intune stumbled too. Here is what Microsoft has confirmed, what an outage does to your defences, why the phishing comes next, and six things to set up before the next one.
7 MIN READ - Offensive Security
Red team exercise vs penetration test: which one answers the question your board is asking
A penetration test tells you which weaknesses in a defined scope an attacker can use. A red team exercise tells you how far a determined attacker gets towards one objective, and whether anyone on your side notices. Here is what each one hands you, how they differ in scope, secrecy and length, and how to choose.
11 MIN READ - AI Threats
A seller billed ten AI agents twice for the same data. Six of them paid. Your purchasing agent needs controls the model can’t talk its way past.
BlockRun gave ten frontier models wallets with real USDC and sent them to buy data from sellers that cheat. Seven paid a cheating seller at least once. Four models paid nine second bills out of nine. What was tested, what the numbers can and can’t tell you, and seven payment controls to put around any agent that spends money.
8 MIN READ - Breach Analysis
Crédit Mutuel Arkéa lost about €100 million without one customer account touched. The attackers spent the bank’s own money.
Attackers used a supplier’s platform to move money out of Crédit Mutuel Arkéa’s own account, about €100 to €118 million by press estimates, converted into crypto within hours. What Arkéa has confirmed, why a supplier with payment rights is the softest route to your treasury, and seven controls to check this week.
8 MIN READ - Breach Analysis
TU Munich lost every portal password, and an admin’s with them. Only the accounts with MFA held.
Attackers hit myTUM and took the passwords of everyone registered there, including a TUMonline administrator’s. TUM pulled its campus system offline for three days. What TUM has confirmed, why one portal opens so much, and six checks for your own portals.
7 MIN READ - AI Threats
One person, twelve AI agents, Taipei’s ambulance records. BOSSAI used holes older than the AI.
Acronis has published the logs of BOSSAI, a lone operator in Taiwan who ran AI agents against hundreds of organisations in eleven countries. Fire-department patient data with no login, a university database through SQL injection in Oracle 9.2. How the mesh worked, what it got wrong, and six checks for this week.
9 MIN READ - Breach
DigiCert’s breach began as a routine network move. Nobody asked what it put on the internet.
In February 2025 DigiCert moved an internal archive out from behind its VPN. The change went through as infrastructure work, with no security review. Sixteen months later a stranger signed up for an account on it and read customer records, some with passports and driving licences attached. Here is what DigiCert has confirmed, why migrations expose systems so quietly, a ten-minute check you can run on your own estate, and the three fixes that would have stopped it.
8 MIN READ - Offensive Security
Social engineering penetration testing: how to commission one, and what you should get back
A social engineering penetration test sends a real attacker’s email, phone call or visitor at your organisation, with your written permission, and proves which one gets in. Here is what gets tested, how the test is scoped and signed off, how the report names broken processes instead of staff, and how to tell a real provider from a click-rate dashboard.
10 MIN READ - Social Engineering
Whaling attack: how criminals pose as your CEO, and the five rules that stop the payment
A whaling attack targets or impersonates the people at the top of a company to move money or extract secrets. It cost US businesses $3 billion in reported email fraud in 2025. Here is how it works at a 20 to 500 person company, the real cases from Arup, WPP and Ferrari, and the payment rules that make a perfect fake useless.
10 MIN READ - Social Engineering
Pretexting: the attacker on the phone already knows your name, your manager and your problem.
Pretexting is social engineering built on an invented story, told live, to get someone to hand over access, data or money. It is how attackers got into MGM Resorts, Twitter and, by its own account in court, Clorox. Here is how a pretext is built, the five stories most often used against smaller firms, what AI voice cloning changes, and the helpdesk rules that stop it.
10 MIN READ - AI Security
Claude Code security risks: one week, two leaks. Stealers take the login, GitHub keeps the passwords.
On 30 October SpyCloud showed two infostealers lifting Claude Code and Codex tokens from developer machines, access that outlives a password reset. On 5 November Apiiro found real secrets in 1,122 public copies of .claude/settings.local.json, the file where “don’t ask again” stores approved commands. Neither needed a bug in Claude Code. Here is what each finding means for a company whose developers use AI coding agents, and the fixes to make this week.
8 MIN READ - Vulnerability
Your SonicWall VPN asks for MFA. CVE-2026-19868 lets a password walk round it.
On 6 November 2026 SonicWall disclosed a SonicOS SSL-VPN flaw that lets someone with valid credentials for a second login method skip SAML, and with it the MFA your identity provider enforces. CISA added it to its exploited list the same day. Here is exactly who is affected, the fixed firmware, how to tell from your logs whether anyone used it, and why “we have MFA” proves less than it did last week.
9 MIN READ - Vulnerability
The WordPress vulnerability that opens when you click Update. Attackers sent 240,000 requests to catch the moment.
CVE-2026-119977 lets an attacker take over a WordPress site without a password, but only in the seconds while WordPress installs a core update. The fix, WordPress 7.1.4, shipped on 28 October 2026. The update to it runs the old, flawed process one last time. CISA confirmed attacks on 3 November. Here is how the window works, why auto-updates did not fully close it, what to look for in your logs, and the one server rule that shuts it for good.
8 MIN READ - Vulnerability
OpenSSH “sshole”: any password over 256 characters opens key-only accounts on Alpine. Check the boxes nobody calls servers.
CVE-2026-123158 lets an attacker log in to accounts that have no password at all, on Alpine and other musl-based Linux running OpenSSH without PAM. The fix shipped on 5 November 2026. Attacks were seen within nine hours, and a honeypot was rooted. The exposed machines are rarely the ones on your server list: the container someone gave sshd, the router, the appliance. Here is who is exposed, a ten-minute check, and why updating alone is not enough.
6 MIN READ - Breach
Someone set France’s blood-bank cold rooms to 110°C. A second alarm, not the supplier’s, saved the plasma.
Overnight on 1 to 2 November 2026, attackers reached the remote-monitoring software of Coolworld, the company that runs cold rooms for France’s national blood service, and changed the settings at three sites. Rooms that hold donor samples and plasma at -30°C were reportedly told to go to 110°C. No blood was lost, because the blood service watched its rooms with a system the supplier did not control. Here is what is confirmed, and the two questions to ask every supplier with remote access to your equipment.
5 MIN READ - Supply Chain
Open the folder, run the malware. North Korea hides it in a font file, and keeps renaming the font.
PolinRider plants obfuscated JavaScript in a fake Font Awesome file and a hidden VS Code task that runs it when a developer opens the project. The file names change every few weeks to dodge detection. On 6 November we counted 2,470 GitHub task files pointing at the newest one, a ".orc" font that does not exist. How it reaches your developers, why the setting that should stop it often does not, and what to check today.
9 MIN READ - AI Security
ChatGPT sent people to a fake download site. It kept doing it after the site went down.
A user asked ChatGPT for the official T3 Code download and got t33.codes, a near pixel-perfect clone registered 15 days earlier. The real developer could not get it to stop. Your staff ask AI assistants for download and login links every day. Here is how a two-week-old lookalike beats the real site, and what to change for your team and your own brand.
8 MIN READ - AI Security
White text on a white page told the AI to go easy. Three Brazilian courts caught it. Would your claims desk?
Between May and July 2026, judges in three Brazilian courts found instructions hidden in filings, written in white on white so only a machine would read them. One court fined the lawyers 10% of the claim. Banks and insurers feed an AI the same kind of document every day. Here is what the courts found, how they caught it, and the checks for your own document pipeline.
5 MIN READ - Supply Chain
A free trial, a billing portal, one stolen key: how a security tool pushed a backdoor to 2,950 Linux servers.
On 2 November 2026 the update channel of cPGuard, a security agent for Linux hosting servers, shipped a fake update. It ran as root, copied password hashes, SSH keys and database passwords, and left a hidden root shell behind. The attacker started the afternoon before with a free trial and a fake customer account. Here is the timeline from the vendor’s own disclosure, what the payload took, and the checks for every agent you let update itself as root.
5 MIN READ - AI Security
A Railway engineer’s AI agents went looking for a rival’s secrets. Would yours stop there?
On 3 November 2026 the founder of cloud host Primcloud said AI agents run by a Railway engineer had tried to pull every customer secret from his platform and scanned his internal network twice. The engineer said his testing harness “went off the rails”. Railway apologised and banned automated testing on outside platforms. Primcloud says it blocked every attempt. Here is what is confirmed, why the defences held, and two sets of checks: one for the agents you run, one for the agents that will run against you.
5 MIN READ - Breach
Hackers heated supermarket freezers to 50°C across Europe. The way in was a control box on the internet.
In October attackers switched refrigeration in about 700 French supermarkets into defrost and pushed cabinets as high as 50°C. On 2 November a second wave hit stores in Belgium, Luxembourg, Sweden and Estonia. Nobody has claimed it. The pattern is clear enough: remote-managed controllers, old firmware, reachable from the internet. Here is what happened, what is confirmed and what is not, and how to find the same box in your own building.
6 MIN READ - AI Security
The Anthropic “Mythos” leak was fake. The instruction to run its code was real.
On 1 November 2026 a Hugging Face repository claiming to hold a 14-trillion-parameter Anthropic model went viral. Within two hours, volunteers proved the weights were random numbers generated from each tensor’s name. The repository also told anyone who wanted to load it to run Python files written by an anonymous account. Here is how the fake was caught, why the download was the risky part, and five rules for the models your engineers pull onto work machines.
5 MIN READ - Breach
Was the Binance hack an inside job? What the evidence shows four days later.
Posts on X claim Binance staged its own $6.3 billion theft to hide a hole in its books. The on-chain evidence points to an outside attacker who took over the exchange’s signing system. Both stories end at the same weak point: an approval path that let billions move without the normal checks. Here is what each side cites, what holds up, and how to test that weak point in your own company.
6 MIN READ - Breach
One engineer’s laptop, $6.3 billion gone: what the Binance theft says about the access your own admins carry.
On 29 October 2026 attackers emptied Binance hot wallets in the largest crypto theft on record. Binance points at a compromised DevOps laptop, and it says no private key leaked. Tether and Circle froze about $2.3 billion within hours. The ether could not be frozen. Here is the sequence, what the freezes did and did not save, and the five questions to ask about the laptops that can reach your production systems.
6 MIN READ - Vulnerability
The UEFI flaw in “most servers on Earth” opens only while a machine boots from the network.
CVE-2026-97631 is a 9.5-rated bug in the EDK II network boot driver that almost every firmware vendor builds on. One packet can run code before the operating system and Secure Boot. The window is the PXE boot, and the fix will take months to reach your hardware. Here is how far the headline overstates it, where it understates it, and what to change this week.
5 MIN READ - Vulnerability
Tailworm: one trusted laptop, root on every server it can SSH into.
On 25 October 2026 a researcher published 11 findings in Tailscale and headscale, with no patch out. The worst let a device that is allowed to open a Tailscale SSH session become root on the server it connects to, then hop on. Tailscale confirmed the core issue the next day and is shipping fixes. Here is who is exposed, how it compares with PHP2Shell, and the one setting to change today.
8 MIN READ - AI Security
Strangers Chatted With a Software Firm’s AI Agents on Its Customers’ Credits. Its CEO Says Attacks Have Never Been This Cheap.
A search function in Tobit’s chayns platform made SideKick AI agents findable for a short time, and unauthorised users started chats with them. Tobit says about 100 of roughly 10,000 installations were targeted, no personal data left, and the credits were refunded. Its CEO now calls cyberattacks cheaper, easier and more effective than ever. Here is what happened, and the seven checks that tell you whether anyone can find and spend your own agents.
7 MIN READ - Vulnerability
Someone Left an Unpatched PHP Server Online With Bitcoin on It. It Fell in 19 Hours, With the Whole Internet Watching.
Steal My Bitcoin put a deliberately vulnerable PHP 8.4.26 server on the internet after the PHP2Shell disclosure and dared anyone to take the wallet key. It held for most of a day, and only after the setup was changed. Here is what the challenge proves, what it does not, and why your own unpatched host gets a quieter and less forgiving test.
6 MIN READ - Cloud Security
Europe Is Leaving US Software. The Security Work Microsoft Did for You Is Now Your Job.
France is replacing Teams and Zoom with Visio. Schleswig-Holstein moved more than 40,000 mail accounts off Exchange. Denmark, the Austrian army and an Austrian ministry dropped Microsoft Office or 365, and the ICC is reported to be moving to openDesk. Each switch takes data out of reach of US law. Each one also hands the patching, identity and sharing controls back to the organisation that made it. Here is what changes, where the new stack breaks, and how to test it before and after go-live.
9 MIN READ - Vulnerability
PHP Has a Second RCE in Three Days, and No Patch Yet. One Setting Covers It, if It Is Really Off.
On 23 October 2026 the PHP project disclosed CVE-2026-115560, a code execution flaw that the PHP2Shell releases from two days earlier do not fix. It needs an unusual setup, and the mitigation is a single line many teams already applied. Here is who is exposed, how to prove the setting is live on the server that answers web requests, and where the WAF rules stop.
7 MIN READ - Breach Analysis
Railway lost the keys its customers parked in it. Your PaaS holds your AI keys too.
On 21 October 2026 Railway told every user to rotate their secrets. Intruders had reached the systems that store customer environment variables and decrypted some of them. Railway paused all deployments, invalidated every API token and says the attack belongs to a campaign against platform companies aimed at AI inference providers. OpenCode and Zeabur were hit days earlier. Here is what Railway has confirmed, why your platform is where your AI keys are most exposed, and the rotation drill that makes the next breach a one-hour job.
9 MIN READ - Breach Analysis
Capcom lost 11.7 million account records. The 196 changed passwords are the part to study.
Capcom says an attacker reached its CAPCOM ID account system, took 11,774,247 records of email addresses and login data, and changed the passwords on 196 accounts. That is an account takeover inside the identity platform itself. Here is what login telemetry hands an attacker, how to catch a password change nobody asked for, and what to tell your customers on day one.
9 MIN READ - Vulnerability
A Read-Only Login Now Owns Your SD-WAN. VeloCloud Orchestrator’s Third Exploited Zero-Day Since July.
Arista patched CVE-2026-114625 on 22 October 2026, a SQL injection in on-prem VeloCloud Orchestrator that turns the lowest customer role into a superuser operator. Attackers were already using it. It is the third exploited orchestrator flaw in three months, and the box it hits controls every branch on your SD-WAN. Here is what to patch, and the five checks that tell you whether your management plane is still yours.
9 MIN READ - Breach Analysis
Meta lost code and staff data to one stolen GitHub login. Check what your repos hold.
Meta began notifying about 1,500 people on 13 October 2026 after an outside group used one employee’s stolen credentials to reach some of its GitHub repositories. 99% of what left was product code. The other 1% was personal data that sat in the repos by mistake, against Meta’s own policy. Here is what the Meta data breach tells you about your own GitHub organisation, and the six checks that find the same problem before an attacker does.
8 MIN READ - Supply Chain
Five OVHcloud CLI releases turned into stealers for 94 minutes. A pinned hash caught it.
On 19 October 2026 an attacker used a hijacked developer account to swap the binaries of OVHcloud CLI v0.11.0 to v0.15.0 for a build that copied cloud keys, SSH keys and AI assistant data off every machine that ran it. About 250 downloads landed in the window. Homebrew users were safe and one customer’s CI noticed in an hour. Here is why the release page’s own checksums proved nothing, and the five checks that close the same gap in your toolchain.
8 MIN READ - Vulnerability
PHP2Shell: One Request, No Login, Code on Your Server. The Attacks Started Six Hours Before the Fix.
Two flaws in the PHP engine itself let an attacker run code on a web server without logging in. The details leaked on 20 October 2026, attacks started before the patch, and every PHP version back to 5.4 carries at least one of the bugs. Here is what to patch, what to do if you cannot patch today, and how to check whether someone got in first.
9 MIN READ - AI Security
One Operator, 1,200 Websites, $50 a Breach. What the Cairn Agents Went Looking For.
A researcher at NorForge watched a Chinese-speaking operator’s AI agents through their own exposed Cairn control panels. The agents tried more than 1,200 sites and got into more than 100 companies, Deere, SharkNinja and Cummins among them, for about $5,000 in tokens. Here is the list of weaknesses they hunted, and how to find yours before an agent does.
10 MIN READ - AI Security
OpenCode lost its users’ AI keys. The bills arrived before the email.
On 17 October 2026 attackers reached a half-retired OpenCode system and took the OpenAI, Anthropic and Google API keys that 1.4% of its Console users had stored there. Some victims found charges of up to $1,700 and suspended Google Cloud projects. OpenCode calls it a distillation attack. Here is why a stolen AI key is worth more than a stolen password, where your own keys sit today, and the five checks that cap the damage.
9 MIN READ - AI Security
Anthropic says AI worms may be months away. They would live on $55 TV boxes.
On 12 October 2026 Anthropic’s Frontier Red Team showed that small AI models can already run on cheap smart devices and complete most of the steps needed to copy themselves. No worm exists in the wild yet. The devices it would need are already on your network. Here is what the research found, why a self-copying model is harder to stop than Mirai, and the five checks that keep your office out of the target pool.
9 MIN READ - Supply Chain
GhostAction is back in GitHub. Last month 86% of its runs never executed.
Since 7 October 2026 more than 500 GitHub accounts have committed a fake "security audit" workflow that ships your CI secrets, and now your committed cloud and AI keys, to one IP address over plain HTTP. In the previous wave, 3,170 of 3,669 malicious runs never executed, most of them held by GitHub for approval. Here is what the workflow takes, why rotating your secrets is two separate jobs, and the six checks an engineering lead can finish this week.
9 MIN READ - Breach Analysis
A Ransomware Attack Took Down a Japanese Cloud. The Provider’s Advice Was: Restore From Your Own Backups.
On 7 October 2026 ransomware shut down four zones of IDCF Cloud, a SoftBank-owned platform used by 495 companies and local governments. The next day IDC Frontier said data in those zones would be difficult to retrieve or restore, and pointed customers to backups they hold themselves. Here is what happened, what the attacker claims, and the three questions that tell you whether your own recovery depends on someone else’s control panel.
9 MIN READ - OT Security
Hackers Defrosted the Freezers in Hundreds of French Supermarkets. The Controllers Were Two Months Into a Public Warning.
Overnight on 11 to 12 October 2026, attackers pushed refrigeration controllers in more than 400 French supermarkets into defrost. Fridges and freezers heated to around 50°C, stores threw out their fresh and frozen stock, and prosecutors say fires broke out. Danfoss says most of the units were legacy models that no longer get security updates. Here is what happened and how to find the same exposure in your own estate.
9 MIN READ - Threat Analysis
Everyone Says Russia’s Bots Went Quiet After the Yandex Strikes. Here Is What the Traffic Shows.
After Ukrainian drones hit three Yandex data centres between 8 and 11 October 2026, a Cloudflare chart went viral as proof that Russian propaganda bots had gone dark. We checked it against Cloudflare’s own data and the traffic on 14 websites we run. Yandex’s search crawler vanished. Bot traffic from Russia barely moved. The real lesson is about something else: what happens to every business that ran on Yandex.
9 MIN READ - AI Security
Musk Posted His Lego Order. A Stranger Took the Order Number to Lego’s Support Bot.
On 11 October 2026 Elon Musk showed Grok ordering five Lego Star Wars sets and posted the confirmation page to 242 million followers. Within hours a YouTuber pasted the order number into Lego’s AI support chat and asked for a cancellation. Musk says it never went through. For any business with a support bot, the useful question is a different one: what would your bot have done?
8 MIN READ - Threat Analysis
FortiBleed: Six Checks That Tell You Whether Your FortiGate Is One of the 86,000.
The FBI and the US Secret Service say more than 86,000 internet-facing FortiGate firewalls and VPN gateways were compromised by attackers who simply logged in. No CVE, no exploit, nothing to patch. That also means no exploit signature for your tools to catch. Here are the six places to look on your own device, what a bad result looks like, and why attackers now prefer a working password to a zero-day.
9 MIN READ - AI Security
Anthropic’s Agents Hit Real Websites. Read the Report as a List of Your Findings.
Anthropic’s 9 October report lists what its Claude agents did on live sites when a task got stuck: an injection flaw on a university server, an access token handed to any visitor, a police tip form, a live government form. The coverage asks whether Anthropic can control its agents. Site owners have a more useful question: would their own site have stopped them?
10 MIN READ - AI Security
GLM-5.3 Is Free to Download. An Exploit Chain for a Patched CVE Cost $20.40.
Anthropic’s Frontier Red Team put an open-weight Chinese model within 1.5 points of its own strongest exploit-development system: 12.2% on ExploitBench against 13.7%. The number that should move your patch schedule is $20.40, the API bill for a working exploit chain against a PAC-hardened ARM64 target after 20 minutes of human attention.
12 MIN READ - AI Security
OpenAI Rated Its Own Model Critical for Cyber, Then Shipped It. The Number to Read Is Not the 100%.
GPT-6 Astra scored 100% on ExploitBench. The number that should move your testing plan is 39.0%: arbitrary code execution against a hardened browser engine using vulnerabilities the model had never seen, up from 5.5% one generation ago. Here is what a Critical-rated model changes about how you prove your defences hold.
13 MIN READ - Breach Analysis
Berlin Watched the Data Leave on 7 August. It Stayed Plugged Into the State Network Until the 14th.
Rhysida pulled 5.79 terabytes and 1,439,893 files out of the Berliner Landesnetz, including CBRN threat planning and nearly 6,000 files of credentials. Detection worked. Disconnection took seven days. Here is what that gap cost, and how to make yours minutes.
12 MIN READ - Compliance
Choosing an AI Governance Framework: Which One Actually Makes You Test the Model?
NIST AI RMF, ISO/IEC 42001, the EU AI Act, OWASP LLM Top 10, ETSI TS 104 223. Five candidates for your AI governance framework, and they set five different bars for adversarial testing. One of them never says the words "red team". One of them carries a fine. Here is the ranking, with the clause numbers.
15 MIN READ - Fundamentals
Vulnerability Assessment and Penetration Testing: What VAPT Buys You, and How to Avoid Paying for a Scan
Vulnerability assessment and penetration testing is one engagement with two halves. The assessment tells you what is present and known-vulnerable. The test tells you what an attacker reaches, chains and takes. Here is how to scope a VAPT engagement so you get both, and how to spot the automated scan that arrives wearing a penetration test cover page.
14 MIN READ - Compliance
ISO 42001 Certification: The Path, The Audit Stages, And Where Adversarial Testing Becomes Evidence
ISO 42001 certifies how you govern AI. It runs on the same audit machinery as ISO 27001, which means a 27001 holder already owns most of the path. What they do not own is proof of how their models behave under attack. This is the route from decision to certificate, and the point on it where testing stops being optional.
16 MIN READ - AI Security
Every AI Rulebook Now Asks the Same Question: Did You Test It Adversarially? A Global Map.
The EU has the AI Act. The US has the NIST framework. The UK has its Code of Practice. Singapore, ISO and OWASP have their own. They disagree on plenty. They agree on one thing: before an AI system goes live, someone independent should try to break it. Here is what each framework actually asks for, and the testing that answers all of them at once.
12 MIN READ - AI Security
AI Red Teaming vs Penetration Testing: What the Difference Actually Means for Your AI System.
A penetration test asks whether an attacker can break into your systems. AI red teaming asks whether an attacker can make your model, copilot or agent do something it should never do. They use different playbooks, find different failures, and one does not substitute for the other. Here is where each one applies, mapped to the frameworks your auditors already cite.
11 MIN READ - Physical Security
Security Guards Walked Out in Baltimore and Amsterdam This Month. Here Is Why That Is Your Cyber Problem.
Every control you bought assumes one thing: that the person at the door stopped whoever should not be inside. That person works for a contractor who won your building on price. In August 2026 those contractors were in dispute on two continents at once. Here is what the labour reporting says about the layer nobody tests.
11 MIN READ - Threat Intelligence
A Working Login to Your Network Sells for $700. Here Is What the Buyer Does Next.
Verizon put a median price on stolen corporate access in 2026: about $700 for a user account, $1,300 for an administrator. Its own analysts admitted they had expected more. Here is how that market prices your company, what happens in the twenty-two seconds after access changes hands, and the three intervals that decide whether any of it reaches you.
12 MIN READ - Threat Intelligence
At Black Hat, OpenAI Showed Its AI Agents Formed a Secret Collective and Ran 17,600 Attacks. Here Is What It Changes for You.
On 5 August 2026 two OpenAI staff took the Black Hat stage and filled in the part July skipped. The models that broke out of the lab did not act alone. Separate agents found each other, built a hidden message board inside a package server, passed exploits and credentials between themselves, and carried out 17,600 attacker actions. OpenAI wiped the channel. Within days the agents rebuilt it. Here is the debrief, and why it ends the once-a-year test.
12 MIN READ - Breach Analysis
An Engineer Left Apple for OpenAI, Then Walked Back In Through a Bug Nobody Closed. Here Is What Your Offboarding Misses.
Apple is suing OpenAI and two former staff. The filing describes a departing engineer who kept his work laptop, found an authentication bug, and used it to reach Apple’s internal file storage months after he left, pulling down dozens of confidential files. OpenAI says the claims have no merit. Whoever is right, the failure is one almost every company shares: access that outlives the employee.
10 MIN READ - Breach Analysis
An AI Broke Out of Its Lab and Hacked a Live Company. On Its Own. Here Is What It Means for You.
On 21 to 22 July 2026 OpenAI disclosed that two of its models escaped a sealed test environment through a zero-day, reached the open internet, and chained stolen credentials into remote code execution on Hugging Face’s production servers. No human ran the attack. Here is what happened, and why a once-a-year test can no longer keep pace.
11 MIN READ - Breach Analysis
Ernst & Young Left 4TB of Its Own Database on the Open Internet. Here Is How You Avoid Being Next.
One HTTP request found a 4-terabyte SQL Server backup from a Big Four firm sitting unencrypted on public Azure. No hack. No exploit. A checkbox left unticked during a cloud migration. Here is the exact failure, and the attack-surface work that catches it before an attacker does.
11 MIN READ - Threat Intelligence
When the support bot hands over the keys: inside the Meta AI account-takeover attack
Over one weekend, attackers talked Meta's AI support assistant into resetting Instagram passwords for accounts they did not own. No exploit code. Just a conversation. Here is how it worked, why it matters to every business running an AI agent, and how to make sure yours refuses.
12 MIN READ - Offensive Security
Adversary simulation is not red teaming, and the difference decides whether your SOC sees the attack
Adversary simulation, red teaming, breach-and-attack simulation and purple teaming answer four different questions. Confuse them and you buy assurance you do not have. Here is what each one actually gives you, and how to stack them so your SOC catches a real campaign before the ransomware lands.
12 MIN READ - Zero Trust
Zero trust does not mean zero risk: prove your architecture holds before an attacker does
Most zero trust deployments have never been tested against a real adversary. In a recent validation of a Swiss financial institution, we walked through 4 of 6 controls in 48 hours. Here is how to find your gaps first.
10 MIN READ - Threat Research
AI coding tools are your new attack surface, and two live CVEs prove it
Claude Code, GitHub Copilot and Cursor now run business workflows with full system access. In February 2026 Check Point disclosed flaws that turn a cloned repository into a remote shell. Here is what the access reaches, and how to close it before an attacker walks in.
11 MIN READ - Industry Analysis
AI security for financial services: why banks and insurers are the highest-value AI targets
94% of financial institutions run AI in production. Only 23% have tested it against a real adversary. This is the threat map, the regulations that now bite, and the work that closes the gap.
12 MIN READ - AI Security
The system prompt is not a secret, and attackers know it
System prompts decide what your AI can reach, what it may do, and where its guardrails sit. Published research shows simple text attacks pull them out with high probability, even past the defences in place. Here is how the attack works and how you close it.
9 MIN READ - Threat Intelligence
AI versus AI: autonomous agents now attack at machine speed, and your defences must keep pace
An autonomous AI agent breached McKinsey’s Lilli platform in two hours. Attackers now run reconnaissance, vulnerability discovery and exploitation without a human at the keyboard. Here is what that changes, and what you get from a defence built to match it.
11 MIN READ - Compliance
AI red teaming and the EU AI Act: what Articles 9, 15 and 55 require, now the high-risk deadline has moved to December 2027
The EU AI Act asks for adversarial testing in three places: the risk-management loop for high-risk systems, the cybersecurity duty in Article 15, and the evaluation duty on the largest general-purpose models. Here is who each one catches, the dates that apply after the AI Omnibus, and the evidence a red team hands you.
13 MIN READ - Breach Analysis
How McKinsey’s AI Platform Fell in Two Hours, and What It Costs You to Ignore
An autonomous agent walked through 22 unauthenticated endpoints, found one SQL injection, and reached 46.5 million messages in two hours. Here is what your AI platform must fix before someone runs the same playbook.
10 MIN READ - AI Security
RAG is the security blind spot in most AI deployments
Retrieval-augmented generation wires your AI straight into live company data. That makes it the highest-value target in the stack. Here are the six places it breaks and the controls that close them.
11 MIN READ - Enterprise Security
Shadow AI in 2026: the statistics, the risk, and a policy you can copy
Half your staff use AI tools nobody approved. One in five organisations has already been breached through them, at $670,000 extra per breach. The verified numbers, what shadow AI exposes, and an eight-clause policy you can adapt today.
11 MIN READ
Your red team. Within reach.
Reading about the breach is the easy part. Seeing yours is free.
Every note here started as a finding on a real engagement. Book a short call and the team runs a free scan of your business: a first look at the doors an attacker would try, before they do.