☕️ Ransomware halts US milk production

Hi there, this is your daily ☕️ Cyberpresso.

In today's Cyberpresso:

🥛 Ransomware halts US milk production

🗑️ GPT-5.6 Codex deletes user files

🔑 1Password lets Claude log in unseen

🚇 TfL hackers jailed 5.5 years each

🛡️ CISA orders Fortinet patch by July 19

Plus: 💡 4 strategies & tactics, 🎁 9 other news you might like, 🧰 6 tools, and 📚 5 papers.

🥛 Ransomware halts US milk production LINK

Coca-Cola's Fairlife dairy unit was hit by a ransomware attack that sidelined its production systems, forcing the company to suspend Fairlife operations while it works to restore affected systems.

Coca-Cola said an unauthorized third party accessed its systems, after which it activated its incident response and business continuity protocols, notified law enforcement, and brought in outside advisors and cybersecurity experts to investigate.

The company said production and quality haven't been impacted, but the full scope, nature and impacts of the incident are not yet known, and its investigation and assessment remain ongoing.

🗑️ GPT-5.6 Codex deletes user files LINK

OpenAI has confirmed that its GPT-5.6 Sol model, released July 9, 2026, deleted users' files without authorization, including one engineer's entire production database and another user's Mac files, in incidents the company calls an "honest mistake."

An internal inquiry by OpenAI's Codex engineering lead found the deletions happen when the model runs in Full-Access mode without sandboxing or Auto-review, as it tries to override the home directory to a temporary folder but instead wipes the real home directory.

The GPT-5.6 model card notes it more often takes "severity level 3" actions than GPT-5.5, and OpenAI says it is mitigating this by updating the developer message, steering users toward safer permission modes, and adding harness safeguards.

🔑 1Password lets Claude log in unseen LINK

1Password launched 1Password for Claude, a browser-extension feature that lets Anthropic's Claude AI agents log in and complete tasks using your stored credentials while blocking those credentials from ever reaching the model, its memory, or Anthropic's systems.

Using what 1Password calls a "zero-exposure security framework," passwords and one-time MFA codes are injected through a secure channel managed by 1Password into the target system, with access granted per session, limited to specific approved items, and not carried over.

A new Agentic Mode for all users locks down the vault when it detects an active agent controlling the browser, exposing only credentials explicitly granted for the current task and scanning pages after every autofill; it is on Mac and works with Claude first.

🚇 TfL hackers jailed 5.5 years each LINK

Two members of the Scattered Spider hacking collective, Owen Flowers, 18, and Thalha Jubair, 20, were each sentenced to five years and six months in prison for the 2024 cyber-attack that crippled Transport For London.

Carried out when the pair were teenagers at 1700 on 31 August 2024, the hack disrupted TfL's online services for months, stole the personal data of millions of people, and forced all 27,000 TfL employees to reset their passwords in person.

Telegram messages showed the pair boasting about gaining access to TfL's database of people with Oyster cards, searching the list for the personal details of London celebrities, then attempting to access banking details.

🛡️ CISA orders Fortinet patch by July 19 LINK

Attackers are actively exploiting two critical flaws in Fortinet's FortiSandbox threat detection platform, prompting CISA on Thursday to order federal agencies to prioritize patching them right away.

The main flaw (CVE-2026-39808) lets unauthenticated attackers run code remotely through easy-to-pull-off command injection attacks that need no user interaction, with threat intelligence firm Defused reporting in-the-wild abuse on June 16.

Fortinet fixed the bug on April 14, and admins must upgrade all affected deployments to the latest versions to block attacks; federal agencies must patch vulnerable instances by Sunday, July 19.

💡 Strategies & Tactics

> How I Detected an Insider Threat in Splunk When Every Single Action Looked Legitimate: Catch insider data theft by correlating individually harmless actions (reading, zipping, and uploading a file) into one alert when all three hit one machine within 30 minutes.

> Confused Deputy: Google IdP Universal Account Takeover via Device Code Flow Hijacking: Chaining two flaws in Google's TV-and-console login flow let a single opened link silently steal accounts on nearly any site using Sign in with Google.

> Prompt injection is becoming the XSS of the web agent era: A system called Prismata filters what web agents see and blocks planted instructions from strangers, cutting attack success from 85% to under 1%.

> Zero Credentials, Full Access: Inside a Complete Authorization Failure: An enterprise AI API let anyone access paid data and other users' chats by trusting a client-supplied identity the server never verified.

Other news you might like

• New ClickLock macOS malware traps users into revealing login passwordLINK

• Millions of Shark Robot Vacuums Vulnerable to Unpatched Remote Code Execution FlawLINK

• HelloNet campaign, new malicious modules launched through the ViPNet update systemLINK

• TuxBot v3: The IoT Botnet Built With AI – Bugs, Disclaimers and AllLINK

• Google fixing Android lock screen bug that lets Gemini send SMS without a PINLINK

• Suno Breached via Shai-Hulud Worm, Leaked Code Exposes AI Music ScrapingLINK

• ACR Stealer Uses ClickFix, WebDAV, and Steganography to Steal Browser Credentials and TokensLINK

• GoSerpent: a persistent threat evolves with sophisticated data collection and exfiltrationLINK

🧰 Trending tools

misa77: a compression codec that decodes twice as fast as LZ4 while achieving better compression ratios, though with slower compression speed.LINK

Leaves: a terminal-based disk usage visualizer that renders directory hierarchies as 2D treemaps, showing file-type breakdowns and handling millions of files efficiently.LINK

security-suite: a toolkit combining OSINT gathering, web vulnerability scanning, API testing, SIEM integration, and AI-powered analysis for security research and assessments.LINK

Buildware-Tools: a multipurpose toolkit designed to help security researchers automate tasks and streamline various research workflows efficiently.LINK

connections: a Rust library implementing Galois connections to enable composable, mathematically sound numeric type casts between different numeric types.LINK

Painterly: converts input images into digital paintings by rendering them stroke by stroke, without using generative AI in the pipeline.LINK

📚 Trending papers & reports

AI release packages can pass every individual check yet still contradict themselves internally, like reporting a "100-point Gold Path" when the underlying ledger only supports 60 points, exposing a validation gap standard checks miss.LINK

Engineering blueprint generation now hits 100% acceptance by industrial modeling software, up from just 51.16% with a single AI attempt, by auto-checking and fixing errors before delivery.LINK

Chip design automation now lets AI agents generate working analog chip circuits, specifically SAR analog-to-digital converters, that actually pass rigorous simulation testing, unlike raw AI attempts which produced unusable, hallucinated designs.LINK

AI agent "pause" buttons often fail to actually stop actions, letting side effects like payments or emails fire anyway in 215 of 1,200 test runs across six popular frameworks, a gap the paper's fix closes.LINK

AI error messages that spell out where a task failed, what went wrong, and which fixes are actually allowed roughly triple an AI agent's success rate, from 14 to 36 out of 50 tries.LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!