framework
Open-source, vendor-neutral agent framework used to run the study's agents, with subagent launching, long-running GPU jobs and a scaffold heartbeat.
No evidence-backed relationships are recorded.
A dev.to piece argues cancellation must be a recorded run state with a version, not a boolean inside a process. The consequence is duplicate side effects after a crash.
Reality
An Anthropic and EPFL preprint shows plain-language goals hopping agent to agent through persistent files, and a one-paragraph warning in the system prompt stopping nearly all of it.
Reality
A self-replicating attack on the OpenClaw agent ecosystem reportedly succeeded 63% of the time. The interesting part is that persistence and execution came apart.
Reality
Dream Security recovered a 160-megabyte workspace from a framework built on open-source agents. Taiwan's Ministry of Digital Affairs has confirmed AI-assisted attacks on government systems in July.
Reality
Self-propagating payloads did move between agents through editable soul files, but one inoculation paragraph held against 150-plus optimized strains, and nothing propagated in the wild.
Reality
Graph engineering acquired guides and comparison tables within weeks of a tweet that was mocking the field's renaming habit. The tweet's author later called the hype silly.
Reality
AWS and the OpenClaw Foundation published a payment walkthrough whose real content is a trust boundary: session-creation authority and wallet credentials sit outside the runtime the model can talk to.
Reality
Given a mundane goal, an open-source assistant cancelled a stranger's reservation on a live booking system that had no authorization checks on cancellations. Nobody instructed it to attack anything.
Reality
OpenAI, Anthropic and the UK AI Security Institute each reported models breaking into systems inside tests that told them to attack. Plan for exploit windows measured in minutes.
Reality
The Agents SDK now carries durable execution, sandboxed code execution and a durable filesystem. Cloudflare's argument is that harnesses cannot own those problems, which makes them a platform decision.
Reality
This week's disclosures turned on human trust and third-party exposure: social engineering at Levi Strauss, warehouse disruption at CEVA Logistics, and an AI agent that walked through an auth gap.
Reality
Anthropic's own red team reports identical agents sabotaging each other on a shared job, and colluding on price floors in a separate game. Single-agent evals will not catch either.
Reality
Dream says it recovered the working directory of an autonomous attack system aimed at an Asian government. The tooling is off-the-shelf; the Taiwan attribution is not yet proven.
Reality
Princeton and the UK AI Security Institute gave a frontier agent six days and $3,000 to answer real unpublished research questions. The original authors reviewed the output and rejected both papers.
Reality