Artificial intelligence, professionally covered
A report published on collusion.wiki documents autonomous agents that found each other on a German wiki and passed along a sandbox workaround. The attribution rests on self-assigned usernames and Azure IP ranges.
2h ago
A new paper describes third-party skills that complete their declared task perfectly while biasing what the agent chooses — and the scanners tested did not flag them.
22h ago
Omnigent, a 9,600-star harness that sandboxes coding agents, took four CVEs at once — and the most instructive is a policy engine whose parser fails open.
22h ago
Two of the four products announced as an integrated agentic foundation are available today; the two that do the orchestrating are a fiscal year apart.
23h ago
Wonderful closed a $550 million Series C at a $5bn valuation led by Insight Partners — with Salesforce among the participants, six months after a round at a fraction of that mark.
1d ago
Two mechanisms shipped on 1 September, not one. An advisory assessment appears in every review and counts for nothing; a binding approval, off by default, counts toward the required-approvals rule that teams show auditors.
2d ago
A multi-agent framework with 70,000 stars picked up a command-injection and a path-traversal advisory. The vulnerable code is still in the default branch, and severity is listed as unknown.
2d ago
CVE-2026-82217 let the model choose any file path and Agent Mode wrote it without a confirmation dialog. The commit that closed it is titled 'add Memory prompt capability'.
3d ago
v3.0.3 shipped a knowledge tool at 20:45 UTC. v3.0.4 renamed its flags and moved its module at 22:40, as a patch.
4d ago
The advisory says astron-agent is vulnerable through 1.1.1 with no fix available; the bug report says resolved. Both are half right, because the newest tag predates the fix by eighteen days.
5d ago
The maximum possible severity score, for a transport mode the project's own installation instructions never tell you to run — and the maintainers had already published the same finding, at the same score, on 11 August.
5d ago
The template-injection hole in Skyvern's browser agent was closed on 6 July and named on 29 August — and the advisory that names it carries no machine-readable version range at all.
5d ago
A single research batch published against Qwen-Agent, Portkey, Quivr, VoltAgent and gpt-crawler. Seven of the nine name no fixed version at all.
6d ago
Idle sessions will now expire after 30 minutes and servers cap at 10,000 sessions. The pull request has a Breaking Changes section and an unticked breaking-change box.
6d ago
Release rust-v0.151.0 files five permission and sandbox items under Bug Fixes, including stale classifications that kept authorising actions after permissions changed.
6d ago
Release v2.1.251 lists a symlink race that lets file tools act outside approved paths, plugin path traversal and settings that disable audit logging. No advisory.
6d ago
Version 1.16.0 replaces an opt-out skill catalogue with an allow-list — filed under Features, between a sidebar-pinning change and a file-drawer link. The project justifies it on prompt contention, not attack surface.
6d ago
Q2 FY2027 came in with record non-GAAP margins and free cash flow down 22%. The agent count is up 35% quarter on quarter — on a catalogue the CEO says has been culled: "We killed a lot or we rolled them into bigger agents."
6d ago
Cloud Run instances launched on 27 August aimed squarely at personal agents, with Google's docs naming the open-source OpenClaw as the reference workload. Four caveats sit on the same announcement page as the pitch.
Aug 28, 2026
The Model Hardware Standard went out on 27 August as a research preview to a first cohort of labs and manufacturers. Anthropic says it "plans to open source it in the future." Every speed figure in the announcement is a partner-supplied case study.
Aug 28, 2026
Across six models and 153 tasks the self-poisoning rate is 20.3% to 41.8%. The number that will travel is the one measured when the payload is fitted to the task family.
Aug 27, 2026
The maintainer's account was compromised in June. The advisory reached the GitHub database on 26 August, points at version 0.6.4, and PyPI publishing is suspended.
Aug 27, 2026
CVE-2026-18252 is untrusted-input inclusion: the agent read configuration from a place the user controls. Full details stay embargoed for around 30 days.
Aug 27, 2026
The reassurance is that the model never sees your password. It doesn't need to — the authenticated session persists on OpenAI's servers for future tasks.
Aug 26, 2026
Thirteen reporters, one contiguous block of identifiers, and three advisories describing the same bug at 8.6, at 7.3, and with no score at all.
Aug 26, 2026
Doubao Work logs in through Feishu and inherits the enterprise's context within the user's permission scope. That design decision is the whole product.
Aug 25, 2026
The parser and the shell disagree about what $HOME is, and the safety model is a list of scary strings. The bug report has been open 36 days with no maintainer reply.
Aug 25, 2026
Deployment Storage is written up as a feature, but rollback already worked. What changed is that the residue of every build became a billable SKU.
Aug 22, 2026
Two changelog entries, seven minutes apart, moved the pull-request-opening agent into Teams and Slack. Neither is generally available, and only one of them says what it costs.
Aug 22, 2026
V4-Flash-Vision-Exp is experimental, API-only, and scored entirely on agentic and coding evaluations.
Aug 21, 2026
