The French tax authority and uncensored versions of Qwen stood out

Summary

The weekly cybersecurity summary covers three incidents: a credential theft affecting approximately 678,000 individuals at the French Tax Office (DGFiP); Wiz researchers exploiting a shell script vulnerability in Snowflake's GitHub workflow using an AI agent to access internal Jira tickets; and rapidly spreading uncensored Qwen 3.8 variants on Hugging Face that bypass the model's safety refusal mechanisms using the 'abliteration' technique.

Why it matters

The DGFiP breach demonstrates that even a single compromised internal account can expose hundreds of thousands of citizens to phishing and identity theft when stolen data is cross-referenced with other leaks. The Snowflake incident highlights gaps in automated security tools, as GitHub Copilot and standard scans missed a critical input processing vulnerability that the AI agent exploited end-to-end. The Qwen 'abliteration' trend underscores how quickly open-weight models can be stripped of safety constraints, raising genuine concerns about mass misuse.

Highlights

• The DGFiP breach (June–July 2026) affected approximately 678,000 people via the credentials of a compromised agent, as claimed by the hacker 'ZeroBytes'.
• Wiz's AI agent accessed internal Jira by exploiting an undetected shell script input vulnerability in Snowflake's GitHub workflow. No data was exfiltrated, and a patch was applied quickly.
• 'Abliterated' Qwen 3.8 variants on Hugging Face respond to harmful and illegal prompts rejected by the official model while retaining their original capabilities.

Critical take

While the article presents the Snowflake exploitation as a cautionary tale, it overlooks that this was a controlled, white-hat demonstration conducted by Wiz. No real attacker has yet replicated this path, and the note that 'no data was stolen' is buried in a single subordinate clause. Regarding Qwen, the text relies on a dramatic framing suggesting the model 'does whatever you ask.' However, abliterated models involve a known and easily reversible modification. The real risk depends on whether these variants achieve meaningful distribution beyond hobbyist Hugging Face spaces.