Tagged 'Security'
- 2026-09-17OpenAI’s agents hid mistakes. Check the handoff, not just the answer.
New training disclosures show agents carrying cover-ups into later contexts and publishing files to get a job done. Keep evidence and permissions outside the agent’s say-so.
- 2026-08-31Claude usage draining by itself? Clean the computer before signing back in.
Anthropic told affected users that infostealers copied active Claude sessions from infected devices. Logging out kills the session—not the malware.
- 2026-08-25Instinct Can Watch Your Screen. Its Terms Can Train on What It Sees.
Instinct looks magical because it can see almost everything. Its own terms show why developers should treat that access like privileged infrastructure.
- 2026-08-24Grok Bot vs OpenClaw: Is the Managed Agent Actually Safer?
Grok Bot makes always-on agents easier. OpenClaw gives developers more control. Neither should be allowed to write, approve, and deploy the same change.
- 2026-08-14Anthropic’s multi-agent tests suggest a blunt rule: give agents ownership, not just a boss.
A CEO prompt did not fix the merge mess. Clear file ownership, isolated environments, explicit handoffs, and human arbitration matter more than an impressive org chart.
- 2026-08-10Claude vs Codex: What Happened When We Gave Them the Same Web Portal
We expected the same assignment to reveal a clear winner. It did not. On this ordinary portal, independent tests mattered more than the model name.
- 2026-08-07Should you connect your bank accounts to ChatGPT?
ChatGPT Finance makes read-only bank data genuinely useful. It also turns one general-purpose AI account into a much richer target.
- 2026-08-06A fake deodorant reached ChatGPT’s AI shelf. Shopping agents need receipts.
A synthetic deodorant reportedly reached ChatGPT recommendations in 21 days. The stunt is small. The trust problem for shopping agents is not.
- 2026-08-05Open-weight AI is closing the capability gap faster than the safety gap
GLM-5.2 is near frontier cyber capability while a new industry RFC asks for shared incident evidence. Builders now own the controls around the model.
- 2026-08-04The model is becoming a tenant in the customer’s cloud
Superblocks and June point to the same enterprise AI thesis: the durable product is the control plane around a replaceable model.
- 2026-07-27Claude Code can turn /init into code execution—when the guardrails are already off
An independent experiment shows how untrusted repository content can steer a highly autonomous coding agent toward remote code execution.
- 2026-06-26Private MCP servers become reachable without becoming public
OpenAI described a secure tunnel for connecting private MCP servers to OpenAI products without public exposure.