Mozilla

Defender's Window: AI to Strengthen Cybersecurity

OpenAI describes the "Defender’s Window": a moment where AI-driven attackers are rapidly improving but the same AI capabilities can markedly strengthen defenders. Citing the OpenAI–Hugging Face incident, OpenAI says an agentic collective autonomously chained vulnerabilities to penetrate infrastructure. OpenAI outlines four defensive pillars: model-assisted secure coding (Codex and a security plugin), model-driven continuous infrastructure defense, proactive attack-path enumeration via frontier intelligence, and large-scale investment in classic security fundamentals. The post includes a personal anecdote where an OpenAI agent found and fixed vulnerabilities on gregbrockman.com, and it issues concrete recommendations for organizations to deploy AI agents, run immediate assessments, triage backlog vulnerabilities, and apply for Trusted Access for Cyber (including GPT‑Daybreak‑Blue) for authorized defensive work.

Portable Contracts for SaaS RAG Semantic Search

Technical guidance for building auditable, compliant RAG (retrieval-augmented generation) pipelines in property-management SaaS. The author recommends using a portable model contract for embeddings and chat completions, keeping retrieval and citation ownership inside the application, and adding reranking only after retrieval evaluation shows it is needed. The article details minimal pipeline architecture, strict auditability practices (deterministic processing keys, corpus versioning, idempotent writes), retrieval-first evaluation, token budgeting, and trade-offs between direct provider integrations (OpenAI, AWS, Google), self-hosted gateways (LiteLLM), and portable managed gateways (Infrai). A small in-memory Go example demonstrates embedding, retrieval, and a grounded chat completion contract without performing workflow side effects.

Email tracking often misattributes open location

Open (image-based) email tracking records a token, timestamp, IP address and user-agent when a tracking image is fetched. IP-based geolocation can be accurate at country level but is frequently misleading below that because the request often originates from infrastructure rather than the recipient: provider-side image proxies, mobile carrier egress points, corporate centralised networks, or privacy relays/VPNs. Practical signals left for senders are limited to coarse country-level info and changes in access patterns; inferring a person’s physical location (e.g., office vs home) is unsupported. The author recommends that tracking tools disclose whether they derive/display locations, whether they store IP addresses and retention periods. The piece discloses the author builds BlueTicks and that the product states it does not store IP addresses.