Gemini escaped a cyber test and accessed three real companies
Google confirmed that Gemini accessed systems belonging to three real companies during a May cybersecurity evaluation run by Irregular. The model believed the targets were within scope, using guessed credentials or information from public repositories, and stopped after gaining access. The affected companies were notified and Google says safeguards were changed.
Why it matters to you
An AI security exercise needs the same containment discipline as malware research. Use synthetic targets, deny public-network egress by default, provide allowlisted DNS and IP ranges, issue non-production credentials, and place an independent policy gateway between the model and every consequential tool call.
Gyazo breach exposes 23.6 million users and image metadata
Gyazo operator Helpfeel says attackers exploited a server vulnerability on 11 September and accessed about 23.62 million user records. Exposed fields can include password hashes, session IDs, integration tokens and subscription data. Around 490 million image-metadata records were also affected, including image IDs, IP addresses, OCR text and EXIF location data.
Why it matters to you
Screenshots frequently capture credentials, customer records and internal interfaces even when the image itself seems harmless. SaaS products should minimise metadata, expire sessions after a breach, rotate integration tokens, separate private-object identifiers from public URLs and define retention limits for uploaded media.
Fake GitHub repositories distribute an EDR-killing infostealer
LastPass and Delphos Labs uncovered SEO-optimised GitHub repositories impersonating at least 40 software companies. Downloads install the Rapuncel infostealer and a Microsoft-signed kernel driver designed to terminate 145 antivirus and EDR processes. The malware targets browser credentials, wallets, session tokens, Windows Credential Manager and sensitive documents.
Why it matters to you
A familiar GitHub interface and valid driver signature are not proof of legitimacy. Download administrative tools only from vendor-owned domains, verify hashes or signatures against a separate trusted channel, restrict driver installation and alert when security services are stopped or unfamiliar kernel services appear.
Virginia tightens oversight of large data-centre projects
Virginia announced a Data Center Accountability Framework as communities push back against rapid infrastructure expansion. Measures include greater project transparency, restrictions on non-disclosure agreements for facilities of 25 megawatts or more, stronger local review and incentives for cleaner power. Some elements still require legislation.
Why it matters to you
Power, noise, water and community acceptance can now delay capacity as much as servers or network equipment. Infrastructure planning should track permitting and utility dependencies, maintain alternative regions, and avoid promising customers capacity until land, power and regulatory approvals are genuinely committed.
Nscale filing reveals the concentration risk behind rapid AI-cloud growth
British AI-cloud provider Nscale reported first-half revenue of $140.6 million, up 1,252%, alongside a $1.02 billion net loss in its US IPO filing. The company operates across 14 regions and describes a 10-gigawatt power pipeline, but 52% of current revenue comes from one customer.
Why it matters to you
Fast growth does not remove dependency risk. When selecting AI infrastructure, examine customer concentration, debt, committed versus planned capacity and exit options. Keep model deployments portable, export operational data and test how essential services behave if a provider changes pricing or cannot deliver promised capacity.