Two Codex sandbox escapes could reach a developer's host
Researchers at Accomplish found two ways around OpenAI Codex isolation. Heapjack recovered a trust token from memory shared by trusted and untrusted JavaScript, enabling unsandboxed commands even in read-only mode. Overpatch used attacker-controlled patch paths to widen filesystem permissions. OpenAI fixed both reports within eight days.
Why it matters to you
Update Codex Desktop to build 26.818.21641 or later and the CLI to 0.149.0 or later. Treat every cloned repository as hostile: keep coding agents away from production credentials, Docker sockets and SSH keys, and use a disposable VM for unfamiliar code.
Malicious npm packages move execution from install time to runtime
Checkmarx found nine npm packages in a campaign led by indexed-btree, which mimicked the legitimate sorted-btree library and reached nearly two million weekly downloads. Instead of relying on an install script, the malware activated inside BTree.prototype.set when given a specific key, then fingerprinted the host and used messaging services plus an Ethereum test network for command and control.
Why it matters to you
Package-install controls alone cannot stop code that waits for normal application execution. Review lockfiles and transitive dependencies, monitor runtime process and network behaviour, and rebuild from a trusted environment while rotating exposed secrets if any affected package was installed.
AI infrastructure carries up to $300 billion of guarantee exposure
The Financial Times reports that technology companies have provided residual-value guarantees supporting as much as $300 billion of debt for AI chips and data centres, often through special-purpose vehicles rather than direct balance-sheet borrowing. The arrangements depend partly on future equipment values while accelerating infrastructure construction.
Why it matters to you
AI capacity can depend on financing assumptions as much as power, networking and accelerators. Assess providers' financial durability, distinguish funded capacity from announced projects, avoid unnecessary long prepayments and keep workloads portable if pricing or expansion plans change.
Taiwan starts an advanced-packaging park anchored by TSMC
Taiwan broke ground on the 88.7-hectare Baipu Industrial Park in Kaohsiung, where TSMC plans an advanced-packaging validation laboratory and talent centre expected in late 2029. Packaging is essential for combining the high-performance chips used by Nvidia, AMD and Broadcom, making it a strategic part of the AI supply chain rather than a final assembly step.
Why it matters to you
Server and accelerator availability can be constrained by packaging even when chip fabrication capacity exists. Forecast AI capacity across the complete supply chain, qualify more than one provider or region and include concentrated packaging dependencies in business-continuity reviews.
US proposes an AI-incident notification channel with China
After talks in New York on 20 September, the United States proposed a bilateral mechanism for notifying serious AI-related national-security incidents for the US and Chinese presidents to consider. China's response was not disclosed. Future discussions could cover AI weaponisation, critical-infrastructure protection and the prevention of cyberattacks.
Why it matters to you
The same principle applies at SaaS scale: model failures and provider compromises need predefined escalation paths. Set thresholds for disabling AI features, name technical and executive contacts, preserve evidence, prepare customer notifications and maintain a tested non-AI fallback.