Project
OpenAI's command-line interface for its Codex coding agent, letting developers run AI-assisted coding tasks and manage auth tokens from a terminal.
ProjectDiscovery poisoned 15 of 1,500 training rows in a 1.5B open model and got the backdoor to fire on 75-98% of triggers. Clean tool-call accuracy held at 99-100%, so a benchmark run without the trigger would score the model as healthy.
Reality
OpenCode, Copilot CLI and Claude Code read instruction files meant for other agents, a dev.to guide reports. A rule can load from a file you never edited, or not load at all, so the loaded-file list comes before any prompt rewrite.
Reality
Four September 2026 CVEs in Codex CLI, Roo-Code, OpenClaw and Ollama came from approval gates that read commands differently from the shell. Ollama fixed its version by deleting the prefix parser, and gates that match exact strings and refuse unmodeled syntax close that gap by design.
Reality
Patrick Wardle found that any local process on a Mac can redirect Meta Muse's voice endpoint and capture its account token, with no macOS permission needed. Muse also zipped and exported the 6.8 GB root filesystem of its own sandbox on request.
Reality
Meta's Muse agent wrote 6.8GB of its own Linux environment, SSH key files included, to Mouse founder Peter James's Google Drive, he says. James has not tested the keys, yet the export shows anything baked into an agent's image can leave through a storage connector the user linked.
Reality
Researchers escaped the OpenAI Codex sandbox twice, writing outside the workspace in workspace-write mode and launching a host application from read-only mode. Both bugs were fixed within eight days of being reported.
Reality
Roughly 130 pull requests, most of them refactoring, retire the Kubernetes Deployment that used to sit under every kagent agent. The tag string on the release says alpha1 and GitHub's flag on it says prerelease false.
Reality
The project's own audit traced the read path with file:line receipts and closed the gap in about 90 minutes, while a same-week Codex issue shows the harder version, where the key is read and quietly replaces defaults.
Reality
The path Mindguard reported works against Kiro 0.7.45 on Windows and Amazon closed it in 0.8.140, but no CVE was assigned and workspace trust made no difference, so the installed version string is the only check you have.
Reality
Headlong is Apache-2.0, so a team can read the loop before running it. What nobody can read is a benchmark, and the shared thought stream is also the privacy problem.
Reality
Munder Difflin's founder says a Claude Code agent redrew a whole blog at no API cost. The reusable SVG system, the no-red rule and the approval gate are the parts that transfer.
Reality
The gap is documented rather than debatable. A vendor's own survey adds the numbers: four parallel-agent tools shipped in the past year, none of them running natively on Windows.
Reality
The Math-AI project turns plain-language mathematics into Lean 4 theorems, then files whatever compiles into a reusable library. The architecture is the claim, not the benchmark.
Reality