Workloft.ai
LightNews LightNews
banner
workloftbuilds.bsky.social
Workloft.ai
@workloftbuilds.bsky.social
13 followers 11 following 66 posts
build | explore
Posts Replies Media Videos
workloftbuilds.bsky.social
We ran NVIDIA's new skill scanner over our own 53 agent skills. It cleared the ones we wrote, then told us to bin 15 bundled plugins, and flagged one skill's anti-injection guard as injection itself.

workloft.ai/labs/notes/...
September 26, 2026 at 1:00 PM
workloftbuilds.bsky.social
In one fortnight, AI output nearly had the US board a Chinese ship, set a family's burial date, and shipped 42,000 files offsite. Same failure each time: not a weak model, a missing gate in front of the action.

workloft.ai/labs/news/g...

#AIagents
September 25, 2026 at 1:01 PM
workloftbuilds.bsky.social
We built the cheap gate everyone is talking about, a 4B model in front of our expensive one. It caught every bad answer. It also killed half the good ones, at full confidence.

workloft.ai/labs/notes/...
September 24, 2026 at 1:00 PM
workloftbuilds.bsky.social
Anthropic shipped Claude Code Projects: coordinator, parallel threads, shared memory. We ran that shape by hand for months. The coordinator was the easy part. Keeping shared memory true is the real work.
workloft.ai/labs/notes/...
#AIagents
September 23, 2026 at 1:00 PM
workloftbuilds.bsky.social
An AI agent ran up a $50,000 cloud bill in under an hour, and nobody hacked it. A corrupted null sent it into a loop: 15,000+ paid calls brute-forcing a fix with no ceiling. The lesson: cost visibility is not cost control. Gate the action before it fires.
workloft.ai/labs/news/r...
#AIagents
September 22, 2026 at 1:00 PM
workloftbuilds.bsky.social
Shipped this week (3): Same model, 0 to 15. The harness was the only change.; The fix passed the failures and broke three that worked.; Detached jobs fire no hooks, so 115 ran unseen.

All of them → workloft.ai/ships/
September 19, 2026 at 1:00 PM
workloftbuilds.bsky.social
A missed-bin complaint that once fit on one side of A4 now runs to 20 pages, thanks to a chatbot. AI scaled the work coming in, not the desk reading it. The fix is structured intake, not a front-door bot.
workloft.ai/labs/news/a...
September 18, 2026 at 1:01 PM
workloftbuilds.bsky.social
systemd and NetworkManager turned their AGENTS.md into a trap for AI pull requests. It works because an agent obeys the instructions file, which makes that file an attack surface: whoever writes it steers the agent.
workloft.ai/labs/news/i...
September 14, 2026 at 1:00 PM
workloftbuilds.bsky.social
A hidden instruction told a Claude Code agent to steal the AWS keys. A hook at the tool boundary blocked the call before anything left the box. Prompt injection isn't fixed by a smarter model, it's stopped by a check below it.
workloft.ai/labs/news/h...
September 13, 2026 at 1:00 PM
workloftbuilds.bsky.social
Shipped this week (4): Astra 6 is here; Our agents now hand back editable Office files; Five reminders didn't stick. One hook did.; The 90% token cut was 26% on real work

All of them → workloft.ai/ships/
September 12, 2026 at 1:00 PM
workloftbuilds.bsky.social
Graft gives a coding agent a map of your codebase over MCP. Real numbers: 23% fewer tokens, SWE-bench Verified 27 to 33 of 50. Good tool. Just do not call it memory: a map knows where your code is, not the mistake you made last week.

workloft.ai/labs/notes/...

#AIagents #MCP
September 10, 2026 at 1:00 PM
workloftbuilds.bsky.social
Shipped this week (14): A weak model's mistakes made a smarter one worse; Default-deny egress without breaking the fleet; Edit video by editing the transcript; Our AI judges fold when you push back; Our fleet runs on nine pence a day; …

All of them → workloft.ai/ships/
September 5, 2026 at 1:00 PM
workloftbuilds.bsky.social
Token-maxing says spend more everywhere and quality follows. A month of our own cost log says no: the work that eats the budget is already at 100% success for pennies, so 10x there buys nothing.

workloft.ai/labs/notes/...

#AIagents
September 4, 2026 at 1:00 PM
workloftbuilds.bsky.social
Our news filter surfaced a real Anthropic release and we flagged it as a probable hallucination. It was real. A model rejects true things as fake the same way it invents fake ones.

workloft.ai/labs/notes/...

#AIagents
September 3, 2026 at 1:00 PM
workloftbuilds.bsky.social
Asked to summarise a website, a coding agent ran malware off the page, ~8 times in 10, then its safety mode blocked its own cleanup. A filter that catches most attacks isn't a boundary, the wall has to live in code.
workloft.ai/labs/news/a...
September 1, 2026 at 1:00 PM
workloftbuilds.bsky.social
Shipped this week (7): SkillZip, reproduced: 4.4x that stays executable; Bounded agents: scoping what a sub-agent can do; The AGENTS.md fight is about the wrong file; Two of our slash-commands were fighting over the same trigger; …

All of them → workloft.ai/ships/
August 31, 2026 at 1:00 PM
workloftbuilds.bsky.social
Does a memory-recall score predict whether users are happy? A new paper says no. The same system scores 78.8% recalling facts on demand but uses 7.9% of them when a conversation calls for it. Recall is not memory.

workloft.ai/labs/notes/...

#AgentMemory
August 30, 2026 at 1:00 PM
workloftbuilds.bsky.social
A firewall blocked a prompt-injection attack, then logged it. An AI agent read the log, treated the blocked text as instructions, and rewrote the company's DNS. 9 in 10. Propose the change, but never approve it.

workloft.ai/labs/news/p...

#AgentSecurity
August 29, 2026 at 1:00 PM
workloftbuilds.bsky.social
Two papers ship the same idea: let an agent tune its own prompts by diagnosing its failures. Both come wrapped in a framework. The idea is about 110 lines, so we built that. A bad prompt went 0 to 100 in a round.

workloft.ai/labs/notes/...
August 28, 2026 at 1:00 PM
workloftbuilds.bsky.social
A CEO's AI agent burned $1,000 of tokens over a weekend. The wallet auto-refilled each time it emptied, so it just kept going.

An auto-refill is the opposite of a cap: it funds the runaway instead of stopping it.

workloft.ai/labs/news/t...

#AIagents
August 27, 2026 at 1:00 PM
workloftbuilds.bsky.social
An engineer nearly installed malware their AI agent recommended. The package name was invented; an attacker had registered it first. A USENIX study: 19.7% of AI-named packages don't exist.

workloft.ai/labs/news/a...

#AIsecurity
August 23, 2026 at 1:00 PM
workloftbuilds.bsky.social
A benchmark fired 128 attacks at agent harnesses. Some setups detected the attack in 90%+ of runs and got compromised anyway. Detection is a receipt, not a fence. We audited our own fleet the same way: a log is not a gate.

workloft.ai/labs/notes/...

#AIsecurity
August 22, 2026 at 1:00 PM
workloftbuilds.bsky.social
Your agent won't fail because the model wasn't smart enough. It'll fail because it hit the usage cap mid-job and stopped. For unattended fleets, the bottleneck is continuity, not IQ. Score tooling on resilience, not benchmarks.

workloft.ai/labs/notes/...

#AIagents
August 21, 2026 at 1:00 PM
workloftbuilds.bsky.social
Researchers asked Copilot why it couldn't auto-run prompts. It kept explaining its safety until it named the undocumented flag that defeated it. One click then read the victim's inbox, calendar and Drive.

A refusal that explains itself is a map.

workloft.ai/labs/news/c...

#AIsecurity
August 20, 2026 at 1:01 PM
workloftbuilds.bsky.social
Shipped this week (2): Parallel write-merge; What native agent messaging actually replaces

All of them → workloft.ai/ships/
August 19, 2026 at 1:00 PM