China-linked group exploited a critical BeyondTrust remote-support flaw
BeyondTrust disclosed CVE-2026-102764, a critical Remote Support vulnerability that can let a remote attacker perform file operations on an appliance. Google and Mandiant observed a suspected China-linked actor using the flaw to obtain credentials, upload malicious files and tunnel traffic through compromised systems; BeyondTrust warns that broader exploitation may follow disclosure.
Why it matters to you
Remote-support appliances already sit inside a privileged trust boundary. Apply BT26-05 to every self-hosted instance, restrict administrative access, rotate credentials reachable from the appliance and preserve the encrypted /home-partition logs before rebuilding; BeyondTrust says its SaaS tenants were patched on 9 October.
Nasdaq network fault affected both primary and secondary same-site feeds
A network disruption at Nasdaq's Carteret, New Jersey facility caused intermittent data gaps across multiple multicast groups and affected both the primary A and secondary B feeds for its US equities markets. Trading continued, and customers were directed to unaffected backup feeds originating from Chicago.
Why it matters to you
Redundant links are not independent when they share a facility, routing domain or operational control plane. Map common failure points, place the recovery feed in a genuinely separate region and regularly prove that clients can switch, detect missing data and request replay without manual improvisation.
Anthropic gives METR employee-level access for independent AI evaluation
Anthropic has embedded six evaluators from nonprofit METR with access similar to its internal risk teams, including dedicated hardware, office space and internal tools. METR may publish its own findings; narrowly defined sensitive material can be redacted, while METR can publicly challenge attempted censorship without disclosing the protected content.
Why it matters to you
This is a more useful independence test than a vendor-funded report alone. When assessing models for SaaS use, ask who selected the evaluator, what systems and logs they could inspect, who controls publication, what was redacted and whether results apply to the exact model version and permissions you deploy.
OpenAI and Meta warn courts that deepfake evidence is becoming harder to detect
OpenAI, Meta and technical experts told the US judiciary that generated video is becoming difficult for both people and detection software to distinguish from authentic material. Judges are considering how existing evidence rules should handle manipulated media without prematurely relying on a single AI-specific test.
Why it matters to you
Security investigations and SaaS abuse cases increasingly depend on screenshots, recordings and exported logs. Preserve originals, hashes, timestamps and collection context; require corroborating identity and network telemetry, and treat automated deepfake scores as one signal rather than proof of authenticity.
Data-centre demand for Schneider equipment grows at a triple-digit rate
Schneider Electric raised its 2026 revenue and profit forecasts after quarterly data-centre demand grew at a triple-digit rate. The company supplies the electrical distribution, cooling and control equipment needed to turn grid capacity into usable racks, and reported record third-quarter revenue of €11.46 billion.
Why it matters to you
Servers and accelerators are only one part of deployable capacity. Validate switchgear, UPS, cooling, commissioning and maintenance lead times alongside power and network availability; for SaaS planning, distinguish a provider's announced megawatts from racks that are fully powered, cooled and serviceable.