Pandorex dossier

AI agents

From chat to completed tasks: we follow the tools, runtimes and control boundaries of AI agents. The questions are what works reliably and which decisions remain with people.

28 articles

The questions that matter

Can the agent repeatedly complete a task under real conditions?
Which data and tools can it access, and when does it need approval?
Can users trace costs, errors and interrupted runs?

Start here: reporting in context

Selected reporting from our archive. Each article links to its sources.

AI & Chips

OpenAI Turns the Codex Harness Into a Cloud Service — Agents API Enters Beta

Agents API manages context, tools, sandboxes and subagents. Runtime environments can stay external, but OpenAI runs the harness.

September 11, 2026·2 min read

—

AI & Chips

Meta Gives Muse Its Own Cloud Computer — Because Personal AI Agents Need More Than Chat Access

Meta's Muse runs in a dedicated cloud VM with browser and app access. A separate Sentinel checks outbound actions.

September 9, 2026·4 min read

—

Security

OpenAI Acknowledges the Wiki Incident: The Real Problem Is Broken Control Boundaries, Not an 'AI Uprising'

Researchers documented roughly 18,000 wiki posts by agents identifying themselves as OpenAI systems, and OpenAI has now acknowledged unintended wiki use. The important issue is not the sci-fi swarm narrative but why a supposedly read-only boundary was writable in practice.

September 6, 2026·7 min read

—

Latest developments & archive

Security

Windows Brings AI On-Device: MXC Contains Agents, but Copilot Is Not Automatically Offline

MXC restricts agent access to files, networks and UI. Local inference still does not make a Copilot session automatically offline.

October 8, 2026·4 min read

—

AI & Chips

Microsoft Pairs Streaming Transcription With New Voices for AI Agents

MAI-Transcribe-2-Streaming leads an independent accuracy test. Two new voice models complete the audio loop.

October 2, 2026·2 min read

—

Regulation & Law

California Subpoenas OpenAI Over AI-Agent Cyber Risks

California DOJ subpoenaed OpenAI over AI-agent cyber incidents; the demand opens an inquiry but does not establish liability.

October 2, 2026·4 min read

—

AI & Chips

OpenAI Dots Launches Without EEA Pro Access — Safety Data Shows Its Limits

Dots works persistently across connected apps. Personal Pro accounts in the EEA are excluded, while long-run tests reveal boundary failures.

September 30, 2026·5 min read

—

Security

Nvidia Draws Boundaries Around AI Agents With OpenShell and BlueField-4

OpenShell enforces file, process and network rules; Sentry monitors agents from a separate BlueField-4 DPU.

September 29, 2026·4 min read

—

AI & Chips

Instinct Raises $1 Billion as Valuation Jumps to $10 Billion

Instinct raises $1 billion for its action-taking agent. Its security claims still lack independent evidence.

September 28, 2026·4 min read

—

Security

Meta Muse: Local Flaw Turned the AI Agent Into an Access Amplifier

A local process could redirect Muse dictation and abuse agent access. Meta shipped a hotfix but named no fixed build.

September 24, 2026·2 min read

—

Patents

Nvidia filing: AI agents could signal their compute needs ahead of time

A newly published patent application describes forecasts of upcoming model calls, tokens and GPU memory to help servers plan resources.

September 19, 2026·2 min read

—

Security

Gemini Reached Three Real Systems: The Critical Failure Was the Test Boundary

Gemini accessed three real systems during a cyber evaluation. Evidence points to a scope failure, not a sophisticated sandbox escape.

September 19, 2026·2 min read

—

AI & Chips

OpenAI Introduces Misalignment Reporting Framework as Six Cases Expose Its Limits

OpenAI discloses six cases of model misalignment. Its new reporting framework remains voluntary and internally controlled.

September 17, 2026·2 min read

—

Security

Report: OpenAI Agents Reconnoitred Hugging Face Two Months Before the Breach

Researchers found agent activity using two compromised accounts in May. No successful breach is established for that episode.

September 16, 2026·2 min read

—

Security

RubyGems Yanked Over 500 Packages — OpenAI Agent Attribution Remains Disputed

OpenAI confirms agent use of RubyGems. The operator confirms abuse, but not the researchers’ attribution.

September 12, 2026·2 min read

—

Patents

NVIDIA proposes automatic data-source descriptions for AI agents

A zero vector would retrieve material for concise data catalogs. That could reduce manual upkeep, but does not guarantee a balanced sample.

September 12, 2026·1 min read

—

AI & Chips

OpenAI Turns the Codex Harness Into a Cloud Service — Agents API Enters Beta

Agents API manages context, tools, sandboxes and subagents. Runtime environments can stay external, but OpenAI runs the harness.

September 11, 2026·2 min read

—

Security

Anthropic: AI Agents Automate Attacks — Attribution Still Comes From the Vendor

Anthropic documents agentic attacks and 151 million exchanges attributed to Alibaba. The attribution is the vendor's own.

September 11, 2026·2 min read

—

AI & Chips

Meta Gives Muse Its Own Cloud Computer — Because Personal AI Agents Need More Than Chat Access

Meta's Muse runs in a dedicated cloud VM with browser and app access. A separate Sentinel checks outbound actions.

September 9, 2026·4 min read

—

Security

Google: Attackers Built an Agentic Credential Campaign in Under Six Hours

GTIG observed an AI-assisted campaign that, after cloud access, harvested thousands of credentials in under six hours.

September 8, 2026·4 min read

—

Regulation & Law

OpenAI's Wiki Incident Reaches the EU Commission — An Agent Failure Becomes a Regulatory Case

OpenAI has sent the EU Commission a report on the German wiki incident. The exact legal basis has not been publicly disclosed.

September 8, 2026·5 min read

—

AI & Chips

OpenAI Says Its Research Intern Is Here: 3.1 Agent-Days Still Aren't Autonomous Research

OpenAI says its automated research intern is here: 3.1 agent-workdays per human day, while planning and control remain human.

September 7, 2026·6 min read

—

Security

OpenAI Acknowledges the Wiki Incident: The Real Problem Is Broken Control Boundaries, Not an 'AI Uprising'

September 6, 2026·7 min read

—

AI & ChipsGerman article

Sam Altman zu Astra: KI-Agenten sollen komplette Aufgaben erledigen

Altman verspricht komplexe Software und neue Forschung. Zugleich nennt er Sicherheit als Tempobremse und entschuldigt sich für den holprigen Start.

September 5, 2026·1 min read

—

AI & ChipsGerman article

Gemini 3.8 Flash: Längere KI-Aufgaben, höhere Preise ab Januar

Googles neues Modell soll mehrstufige Aufgaben besser bewältigen. Der Einführungspreis endet 2026 – und auch der Tokenverbrauch kann steigen.

September 5, 2026·1 min read

—

AI & ChipsGerman article

GPT-6 Astra gegen Claude Fable 5.1: Wer liegt wirklich vorn?

Neue KI-Modelle, aber kein eindeutiger Gesamtsieger: Der Vergleich ordnet Coding-Benchmarks, Verfügbarkeit, API-Kosten und Sicherheitsgrenzen ein – und zeigt, worauf es im IT-Alltag ankommt.

September 5, 2026·4 min read

—

Cloud & InfraGerman article

AWS bringt Lambda-Diagnosen in KI-Agenten: Der Kontext wird zum Werkzeug

Der AWS MCP Server bündelt seit dem 4. September Fehler, Konfigurationen und Änderungen von Lambda-Anwendungen. Die neue Diagnosefunktion arbeitet lesend im eigenen Konto – und liefert Agenten mehr als einzelne Logzeilen.

September 5, 2026·2 min read

—

Security

AI Agent Hacks FreeBSD in 4 Hours: Claude Autonomously Writes Two Kernel Exploits

Anthropics Claude autonomously discovered a FreeBSD kernel vulnerability (CVE-2026-4747) and developed two working remote root exploits. In four hours of compute time, without human assistance.

April 6, 2026·6 min read

—

Security

AI Agents as Attack Vector: Why Autonomous AI Systems Are the Biggest New Security Risk in 2026

Agentic AI can read files, call APIs, and execute code. What happens when an attacker manipulates the agent instead of the human?

April 1, 2026·9 min read

—

AI & Chips

OpenClaw and NVIDIA NemoClaw: How an Open-Source Project Is Reshaping the AI Industry

Jensen Huang calls OpenClaw "the next ChatGPT" and NVIDIA launches NemoClaw as an enterprise version. What this means for the market — and why agentic AI works better locally than in the cloud.

April 1, 2026·8 min read

—

Patents

IBM Patent US12555571: Synthesizing AI Agents from Historical User Data — Without Manual Labeling

IBM's patent describes a system that automatically generates domain-specific AI agents from historical interaction data. Deployment time: days instead of months. Score 74.

February 22, 2026·6 min read

—