July 2026 Summaries

Factory joined the Open Secure AI Alliance alongside NVIDIA and other members to advance openly shared tools, models, and research for securing software and AI agents. Its contributions include Droid Shield 2.0, which uses two fine-tuned Qwen-based models to identify unflagged secrets in code context and reduce false positives while keeping detected secret values masked; Factory has released the related adapters, configuration, prompts, and calibration guidance publicly. The company also described its Automated Security Review program, which develops threat models for open-source projects, applies STRIDE and OWASP frameworks, traces data across trust boundaries, validates vulnerabilities, and responsibly reports findings, including a template-injection issue in External Secrets Operator identified as CVE-2026-42876.

Factory natively hosts Zhipu AI’s open-weight GLM models, allowing users to run GLM-5.2 or GLM-5.1 in Droids without managing infrastructure, API keys, inference endpoints, or separate billing. GLM-5.2 can be selected through the interactive `/model` command or specified in headless and CI executions with `-m glm-5.2`, and it supports Off, High, and Max reasoning levels, while GLM-5.1 supports Off and High. Both models carry a 0.55x credit multiplier, positioning them as lower-cost options for high-volume, exploratory, iterative, cleanup, and narrowly scoped coding tasks. Factory supports model-agnostic workflows, enabling users to pin GLM for deterministic runs, let Factory Router choose models automatically, or assign GLM to worker agents in Missions while reserving more capable models for validation.

Factory’s Incident Response feature connects its Droid agent to Slack alert channels from observability platforms such as Sentry, Rootly, Axiom, and Datadog, allowing it to autonomously investigate, triage, and potentially prepare fixes for incidents. The system conducts sessions on a configured persistent Droid computer, uses available repositories, environments, and monitoring tools, and reports findings within the original Slack thread so engineers can continue the response with retained context. It is designed to reduce time spent on noisy or duplicate alerts while creating a persistent runbook that records past root causes and response knowledge to improve future investigations. Teams can customize the automation with their own incident playbooks, connect MCPs and other tooling, choose monitored channels and session visibility, and run sessions through user or service accounts. Currently in preview, Factory says the feature has saved hundreds of on-call engineering hours and aims to help organizations keep incident management from becoming a bottleneck as software delivery becomes more automated.

Factory’s July 2026 update to its Desktop App adds integrated tools for reviewing and refining Droid-generated work, including documents, presentations, spreadsheets, PDFs, websites, and code, without switching applications. Users can preview live sites, inspect file-by-file code diffs, and use design mode to select specific visual elements or file areas and submit contextual change requests. The app also supports inline comments on code changes, enabling a pull-request-style review process, while expanded session filtering, grouping, and organization aim to keep work streamlined. These features are available on macOS and Windows for all Factory plans.

Factory natively hosts Moonshot AI’s open-weight Kimi K2 model family within its Droid platform, allowing users to run the models without managing inference infrastructure, GPUs, or provider keys. Available variants include the coding-oriented Kimi K2.7 Code, Kimi K2.6, and Kimi K2.5, which can be selected in interactive Droid sessions through the model menu or specified in non-interactive commands, CI workflows, and Missions using their model IDs. Each version offers Off and High reasoning settings, with High as the default, while K2.7 Code and K2.6 have a 0.4x credit multiplier and K2.5 has a 0.25x multiplier, though users are advised to verify current pricing. The guide recommends Kimi K2.7 Code for focused coding tasks where a lower-cost open model is useful and highlights Factory’s ability to switch models per task or use Factory Router to select models automatically.

Factory’s Droid Shield 2.0 augments deterministic secret scanning for autonomous code commits with two fine-tuned language models designed to reduce false positives and catch secrets missed by fixed patterns. The risk model reviews suspicious lines that did not trigger the scanner and prioritizes recall, while the downgrade model evaluates scanner hits using masked context to determine whether they are likely false alarms without exposing credential values. Trained and evaluated primarily on public CredData benchmark material, supplemented by hand-labeled unknown samples and aggregate non-identifying production priors, the models use Qwen 3.6 35B A3B LoRA adapters and were tested with repository-level holdouts. Factory reports that the fine-tuned models achieved higher ROC-AUC scores than the base model and competitive or stronger results than GPT-5.5 and Opus 4.8 under selected operating thresholds, while aiming for lower cost and latency. The company acknowledges limitations involving benchmark representativeness, sparse real-secret prevalence, masked-context constraints, LLM-generated labels, threshold calibration, and comparison methods for closed models, and has released the model weights and supporting materials openly while offering Droid Shield 2.0 in research preview.