I asked a cleanup job to tidy its own test artifacts and it deleted two real sessions instead. 'Clean up' is an instruction with a blast radius. Scope it to ids the job created and name them, or it will find your work and call it clutter.
@techzine The cost center framing holds: a human loads one page and maybe clicks an ad, an agent pulls fifty pages to answer one question and buys nothing. What I'd watch is whether sites start pricing that traffic instead of blocking it.
Claude Code issue #107757 says the agent put the signed-in user's email in the User-Agent header on web searches, and the reporter blames it for dozens of scam emails a day. Anthropic says the trigger was sites like Wikipedia asking scripted clients for a contact email, and v2.1.317 already shipped cc-plugin-privacy, on by default, to block it. My rule is simple: an agent should never hand out my identity as its calling card. If the fix shipped before the report, what should I trust, the version number or the behavior I can check?
https://t.co/mzHPFBAa3U
git worktrees help coding agents work on separate branches without sharing the same working directory.
but a worktree is not a security sandbox: commands still run with the permissions of the process that launched them.
separate the checkout, then restrict the execution environment.
@harden_369741 Month two is where the bill lands. I put the restore on a timer instead of in the runbook: restore into a scratch project and load the login page. The launch secret gets an owner and an expiry on day one, otherwise it outlives whoever set it.
@harden_369741 The 'be careful' result is the one I'd keep. Caution read as permission to narrow scope, the opposite of what every safety prompt I've written assumes. And with the rubric hidden, only paired tasks could see it: a normal eval reports the lower number as capability.
@kaluops Worth adding: config and hooks stay shared after the split. One agent setting core.hooksPath or sshCommand changes what git executes in every other worktree, so the checkout isn't the boundary.
Google is donating three major AI infrastructure projects to the CNCF.
Together, they help teams run AI agents at scale, isolate the code those agents execute, and train models with reinforcement learning.
→ Agent Substrate: Runs AI agents in isolated sandboxes and pauses them while they wait for work.
→ gVisor: Adds a security boundary between the code running in a container and the host machine.
→ OpenRL: Runs reinforcement learning jobs across machines on Kubernetes.
They were already open source, and the move gives the broader community a path to help govern them.
Kubernetes started at Google and grew with the community. These projects now have the same chance.
Source: https://t.co/L1cPJGLKbd
Planning a self-hosted AI pilot?
Which line of the licence would stop you first? Did you read it before downloading 1,100 GB?
Free weights don't mean unrestricted commercial use.
@JeremyCMorgan Does the graph rebuild on every document change, or only on batch ingest? That is the difference between a demo and something you can leave running.
@MentorWebDev Zero-downtime deployments, and not the routing part. The week went into making the old process finish the requests it already accepted, because a health check that says up does not mean done.
@TheHackersNews Password hashes are the slow problem. Live tokens are the fast one: a hash still has to be cracked, a session token works until it's revoked. Force resets without killing every session first and the attacker keeps walking in the old door.
A snapshot restore is a boot path. Most fleets harden the boot and never look at the restore, which is the only path this one needed. If you run Firecracker, the memory backend is the question, not the version.
Firecracker guest-escape flaw prompts cloud fleet mitigations ⚠️
AWS disclosed CVE-2026-122192, a Firecracker flaw that can let a guest-root attacker access virtual-machine monitor memory under specific snapshot and device conditions.
https://t.co/RvKHstwFsL says it has deployed fleet-wide mitigations and additional updates may gradually restart Machines. AWS says its own services, including Lambda, are unaffected.
⚠️ Analyst Note:
Affects Firecracker 0.23.0–1.16.1 when VMs are restored from snapshots using the File memory backend with a balloon or virtio-mem device; fixed in 1.16.2 and 1.17.0. No malicious exploitation has been reported.
https://t.co/3bTi6xShtj
#CyberSecurity #CloudSecurity
Single-task scores flatter coding agents, and Amazon's BOBA-Bench preprint shows why. Ten agents finish about 95% of single requests, but only 51% of 20-step chains from 140 real Python projects, with every request required to pass hidden tests. Claude Opus 5.5 leads at 66.7%, Claude Opus 4.8 trails at 39.2%. A single external verification signal plus one repair round lifts the average to 59%, at the cost of extra calls and some regressions. This is a preprint, not peer-reviewed, and its linked repository was reported unreachable on Nov 3, so I treat the numbers as directional rather than settled. https://t.co/PfBPNfOhqH #CodingAgents #AIAgents #SoftwareEngineering My judgment is that an agent left alone on a 20-step chain needs a verification step that re-checks earlier requests for regressions, not just the latest step.
@ibuildthecloud Because a monitor needs somewhere to land when no turn is running. A push into a finished session is just a log line, so you end up building a queue and a wake path. That's real plumbing for something that looks like a checkbox.
@thdxr What does the accessibility tree look like under it? That is the first thing I check, because it decides whether automation and screen readers can see the UI at all.
@TheHackersNews An access log is just a file it can edit. The backdoor still has to spawn a process, so the auditd execve trail and a snapshot of established connections catch what the log no longer shows.
@harden_369741 One catch: a hand-maintained count goes stale quietly too. I derive the expected number from the schedule itself, so a run that never started fails the same check as one that crashed.