The AI Too Dangerous to Release Publicly

The ShinyHunters extortion gang is back — and this time they went after the back office.

Oracle PeopleSoft, an enterprise software suite used by large organisations to manage payroll, HR, finance, and supply chain operations, is being targeted in an active data theft campaign, according to BleepingComputer. ShinyHunters claims to have stolen data from 300 instances across more than 100 organisations, and affected customers are already receiving extortion demands.

The attackers say they are chaining old and zero-day vulnerabilities together to gain access — though they note the method does not work on every system, suggesting configuration or version differences affect their success rate.

What makes this particularly sensitive is the kind of data PeopleSoft holds. This is not a list of email addresses. HR and payroll systems store names, national ID numbers, salary records, and employment histories. If your employer uses PeopleSoft — especially in a university, government body, or large enterprise — your personal data may be in scope.

No CVE identifiers have been confirmed in the source material at the time of writing. ShinyHunters has not disclosed the specific vulnerabilities used publicly.

What you should do: If you work in IT at an organisation running Oracle PeopleSoft, check with your vendor contact for guidance and monitor for extortion-related communications. If you are an employee at a large organisation, watch for phishing attempts that reference your employer or personal employment details — this is exactly the kind of data that fuels targeted scams.

On June 9, Anthropic released Claude Fable 5 to the public — and simultaneously locked its twin, Claude Mythos 5, behind a vetted access programme, according to The Hacker News.

Both models share the same underlying architecture. The difference is a layer of safety classifiers. Fable 5 — the public version — routes sensitive requests involving cybersecurity, biology, and chemistry to a less capable model. Mythos 5 keeps those capabilities fully active, but is available only to approved cyber defenders and critical infrastructure operators through Anthropic's Project Glasswing.

Anthropic describes Mythos 5 as the strongest cybersecurity AI in the world. The practical implication is significant: the same model that can help defenders analyse threats is considered too capable to release without restriction. Both versions are priced at $10 per million input tokens and $50 per million output tokens — less than half the cost of the earlier Mythos Preview.

What FortiSandbox is: FortiSandbox is a security tool made by Fortinet that analyses suspicious files and web content in an isolated environment to detect malware before it reaches a network.

What it is: A command injection flaw (meaning an attacker can slip unauthorised operating system commands into a request the server trusts) in the FortiSandbox web interface allows a completely unauthenticated attacker to run arbitrary commands on the affected system.

Who's at risk: Any organisation running an unpatched version of FortiSandbox, FortiSandbox Cloud, or FortiSandbox PaaS with the web UI exposed. No login required to trigger the flaw.

CVSS: 9.1 — Critical. Patch today.

Root cause: The web UI fails to properly neutralise special characters in HTTP request inputs — a weakness classified as CWE-78 (improper neutralisation of special elements used in OS commands). When the application passes unsanitised user input directly to the underlying operating system, an attacker can inject their own commands alongside the intended ones.

Attack vector: An attacker sends a crafted HTTP request to the FortiSandbox web interface — no credentials, no prior access, no social engineering required. The server processes the malicious input as a legitimate OS command, handing the attacker execution capability on the system. From there, they can move laterally (spreading through connected systems on the same network), exfiltrate (copy out) data, or establish persistent access.

Detection strategies: Look for unexpected outbound connections from FortiSandbox hosts, unusual process spawning in system logs, or HTTP requests containing shell metacharacters (characters like ;, |, or &&) in web server access logs.

Recommended actions:

1. Upgrade FortiSandbox to version 5.0.6 or above; FortiSandbox 4.4.x users should upgrade to 4.4.9 or above.
2. Restrict web UI access to trusted internal IP ranges only — do not expose the management interface to the public internet.
3. Review FortiSandbox logs for anomalous HTTP requests prior to patching.

If you or anyone on your team works with JavaScript or Node.js projects, this one is worth knowing.

GitHub has announced that npm v12 — the tool developers use to install software packages — will no longer automatically run installation scripts from third-party packages, according to BleepingComputer. Currently, when you install a package, it can silently run code on your machine as part of that process. npm v12, expected next month, will require explicit approval before any of that happens.

This matters because supply chain attacks — where attackers hide malicious code inside legitimate-looking packages — have surged. This change will not stop every attack, but it removes the silent automatic execution that makes them so effective.

Action: If your team uses npm, brief your developers on the change before v12 drops. Packages that previously installed silently may require new approval steps.

The U.S. Cybersecurity and Infrastructure Security Agency added three vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog this week, according to The Hacker News. The KEV catalog is a federal watchlist — if your organisation works with the U.S. government, patching listed flaws within CISA's deadlines is mandatory.

The three additions: a Cisco Catalyst SD-WAN Manager command execution flaw (CVE-2026-20245, CVSS 7.8), a Google Chrome memory vulnerability enabling remote code execution (CVE-2026-11645, CVSS 8.8), and an Arista EOS network flaw with no patch planned (CVE-2026-7473, CVSS 6.9). The Chrome flaw affects everyday users — if you have not updated your browser recently, do it now.