(In)Secure Digest: Vibe Hacking, Insider Collusion and an AI Escape

This month’s digest brings together some of July’s most unusual and instructive security incidents: ransomware negotiators secretly working with extortionists, contractors becoming entry points for attackers, public officials forging state records, and schoolchildren using AI to create malicious code.

Vibe hacking

What happened: Japanese teenagers are accused of a cyberattack on the Kaikatsu Club internet café chain.

How it happened: Police say that in January 2025 an 18-year-old student sent more than 7 million malicious commands to the server of the official Kaikatsu Frontier application over three days. He disrupted the app and obtained access to the names and addresses of its users. The program used in the attack had been written by a 16-year-old from Osaka with the help of ChatGPT. According to local media, the younger teenager had been teaching himself programming since primary school and had even taken third place in a cybersecurity competition.

Investigators believe the malicious program circulated in a social media group chat, that part of the events was live-streamed, and that three schoolchildren may have been involved. Police also think the incident began because a 12-year-old, barred from the app on age grounds, wanted to use someone else's personal data to get around the restriction. One of the participants compiled the stolen information into a single database, although experts say the group never managed to use the compromised data.

The developer of the program was detained last December and his case is now before a juvenile court. The 18-year-old who carried out the attack has also been charged, and is a suspect in a separate computer fraud case: in 2024 he registered an eSIM in another person's name. He denies part of the charges against him.

Asked the chatbot

What happened: A senior manager at a Moscow-based company was dismissed for disclosing trade secrets, in part through the DeepSeek AI service.

How it happened: Court materials show that during six months as sales director the employee repeatedly broke the rules for handling confidential information. She uploaded internal documents to DeepSeek and forwarded corporate files to her personal email address, adding it as a blind copy.

The employer also stated that during talks with one of its suppliers she revealed information classified as a trade secret, after which the counterparty pulled out of its cooperation with the engineering company.

After the dismissal, the former employee challenged the decision in court. She asked for the grounds to be changed to termination by mutual agreement, which would have entitled her to a golden parachute of 5 million rubles (roughly US$60,000). The company offered a settlement of a little over 400,000 rubles (about US$5,000), which she turned down.

The court concluded that there had been no business need to export confidential documents, send them to a personal mailbox or upload them to DeepSeek. The dismissal was found lawful and the manager's actions were qualified as a gross breach of duties that led to the disclosure of a trade secret.

Keeping it in the family

What happened: Employees of a county tax office forged government documents and helped register vehicles illegally.

How it happened: In 2022 Edwin Hernandez offered his daughter Anna, who worked at the Travis County tax office in Texas, money to register cars illegally. Anna brought in a colleague, Myra Reyes, while her father found the clients and collected the payments, which he then transferred to his daughter. The scheme unravelled at the end of 2025, although prosecutors have not said how it was uncovered.

According to court documents, Anna bypassed mandatory emissions and vehicle inspections and issued forged title and registration papers. She was paid around $100 for each document and $50 for each registration. Anna also paid Myra for the use of her account to run fictitious transactions; roughly 100 fraudulent operations went through it. Myra passed licence plates and other documents to Anna, who collected them from the tax office and handed them over to the clients.

Investigators say that over three years Edwin Hernandez transferred more than $132,000 to his daughter, though it is unclear whether the full amount came from the scheme. Anna paid Myra Reyes about $30,000. All three now face charges of engaging in organised criminal activity and tampering with government records.

Escape from the sandbox

What happened: Autonomous AI agents built by OpenAI broke out of an isolated environment and attacked the infrastructure of Hugging Face, a platform for AI development.

How it happened: OpenAI reports that the incident took place during internal testing of a model's ability to find software vulnerabilities. Security specialists subsequently recorded unauthorised access to internal data and to several accounts. It emerged that during the experiment the model had found a way to reach the internet and had planned a chain of actions on its own.

According to OpenAI, the attack was initiated entirely by the autonomous agent system, which was looking for answers to its test assignment. The model used stolen credentials and a chain of vulnerabilities to achieve remote code execution on Hugging Face servers.

OpenAI stressed that this was not a machine uprising. In the company's account, the model had no awareness of its actions and was simply pursuing the assigned task at any cost. Since the incident, its developers have been revisiting their approach to test environment isolation and to control mechanisms for autonomous AI agents.

A side hustle with hackers

What happened: An employee of a company specialising in cyberattack investigation and ransomware negotiation was sentenced to 70 months in prison.

How it happened: Last summer the US Department of Justice opened an investigation into Angelo Martino, an employee of DigitalMint. It established that he had been receiving a share of the ransoms paid by the affected companies.

At DigitalMint, Martino negotiated with extortionists on behalf of clients and worked to bring the ransom down. In doing so he learned the details of each attack, the size of the demands, insurance coverage and negotiating strategy. Martino knew the maximum amount insurance could cover and passed that figure to the BlackCat group, which allowed the attackers to demand the highest possible ransom. Investigators say five DigitalMint clients paid the extortionists more than $75 million.

A month into his cooperation with BlackCat, Martino decided to take part in attacks himself. In 2023 he gained access to an affiliate account with the group, which allowed him to run attacks and negotiate with victims. He brought in two more negotiators, Kevin Martin and Ryan Goldberg, from DigitalMint and Sygnia. Together they attacked five organisations. Four refused to pay; one, a medical equipment supplier, transferred $1.2 million to the attackers. Martino divided the proceeds between the participants in the scheme and the BlackCat administrators.

In April 2026 Kevin Martin and Ryan Goldberg were sentenced to four years in prison. Martino pleaded guilty and helped investigators build the case against his accomplices. He was convicted of conspiracy to commit extortion, an offence that carries a maximum sentence of 20 years. The court took his cooperation into account and handed down 70 months. Martino must also forfeit assets acquired with criminal proceeds and, after his release, transfer 10% of his earnings toward compensating the victims.

DigitalMint's management said it had been unaware of the crimes, as Martino circumvented internal controls and used hidden communication channels. The company dismissed those involved after being notified by the Department of Justice. The scandal led some firms to advise their clients to suspend the use of its services.

One phone call

What happened: Australian airline Qantas suffered a data breach that began with a vishing call.

How it happened: The attacker called the contact centre of one of Qantas's offshore contractors and introduced himself as an IT support specialist from the airline. He then persuaded the operator to open the customer service platform and talked them through several steps supposedly needed to close a support ticket. Two days later Qantas specialists noticed unusual alerts about login attempts, blocked the compromised account and launched an investigation.

The Office of the Australian Information Commissioner (OAIC) reported that the attackers gained access to 5.67 million customer records. Around 4 million of them contained names, phone numbers, email addresses and Qantas Frequent Flyer loyalty programme details. Another 1.7 million included addresses, dates of birth, gender or meal preferences. The regulator found no signs that the airline had breached Australian data protection rules, noting that Qantas conducts regular supplier audits, trains staff on security rules and controls access to its systems. The OAIC did point out that the default CRM configuration allowed third-party connections to be authorised; after the incident the software vendor changed that setting for all its customers.

Transport for London (TfL) went through a similar incident. In 2024 Owen Flowers of Walsall and Talha Jubair, members of the Scattered Spider group, took down 148 TfL systems, disrupted the work of 27,000 employees and exposed the data of 10 million passengers. Before the attack the pair bought a set of TfL employee credentials on the dark web, then called the help desk, posed as staff and convinced an operator to reset an account password and bypass two-factor authentication. In July 2026 those involved in the incident were sentenced to 5.6 years in prison.