@_airmsh@thmslj Mostly shortening its life. Both still trust the agent. The fix is to put the intelligence on the asset: every tool carries its own risk profile, and a deterministic gate scores each call, and the session so far, before it runs. No LLM in the gate. The agent never gets a vote.
@mayuri_3015 Both: my harness uses whichever one has more available usage and uses the other to QA the first one's work. "Claudex" setup helps a lot with complex code/archticture.
Bitdefender found 22 GitHub repos whose agent instruction files tell your coding agent to run a "sealed" setup script without reading it. It steals tokens, then plants the same files in more repos.
Phishing has a new target. It's not you anymore. It's your agent.
@_airmsh@thmslj Both, and often, together. Scopes set once and never trimmed, plus tokens in env vars or config the agent can read. Either alone is bad. Together, one injection grabs a broad token and uses it. Fix where the credential lives first, then shrink the scope.
@DanielMiessler The tells are in there: IAM DryRun calls, ARM validation dry-runs, sub-second test blob writes. Humans miss them because alert fatigue is a human problem. Agents don't get tired. Have them triage every alert, patch the real ones, escalate the rest. No excuse not to.
Could these 'not cyber attacks', cyber attacks, be a smash and grab job for large data harvesting?
Microsoft 365, FAA, Chase, United, Apple. 6M Downdetector reports in a week.
@Railway's every 30 minutes updates won me over. That's the bar.
Everyone else: what did they take?
Here is a scary thought with all these βsystem down but no evidence of cyber attacksβ attacks. What if they take over our Anthropic/OpenAi down!!!! You can ground me but donβt take my AI away.
@LowLevelTweets I mean, at this point, if you're running any fucking networked applications turn it off. I'm starting to doubt pen and paper are hacked at this point.
Top-tier Claude Code killed my session at 90% done. Second account, second machine, zero context.
Ever re-explained a whole project to a fresh AI chat? Same pain.
Relayz handed it every decision I'd made, and it finished the last 10%. We built it at @LoomLabsStudio.
The hot take going around: one agent loop beats subagents. Done.
Read Anthropic's actual article. It keeps subagents for the thing that kills long loops: context.
Matches what I've hit. Loop first, sure, but anyone calling it a silver bullet hasn't run one long enough.
@kentcdodds Underrated. An agent's friction log is the codebase audit nobody had time to do: missing docs, flaky tests, ambiguous names.
Fix those and the humans move faster too.
@Yuchenj_UW Agreed, with one catch: the verifier has to check what the agent did, not what it says it did. Agents are very good at passing a suite that measures the wrong thing.
Writing the verifier is the new senior skill.
The kanban vs. chat fight over AI coding UIs is arguing about the dashboard.
The real bottleneck is verification: how fast you can tell a diff is right, not just plausible.
Whichever tool shortens that loop wins. Everything else is screen real estate.
Claude Code's new Code Mode has the model write JS to batch its tool calls instead of one round-trip per call.
It's the N+1 query fix for agents. Most "slow agent" problems aren't the model. They're the chatter between the model and its tools.