Topic
Security measures that limit an AI agent's actions, often by monitoring or restricting its tool calls, to prevent harmful or unauthorized behavior.
Coinbase says a Qwen3.5-9B model post-trained for under $100 beat Claude Opus 4.5 on its Onramp fraud benchmark. The figures are Coinbase's on private data; the reusable parts are the label-based reward and the tier-only, block-only harness.
Reality
Pink's 80-run test of Claude and Gemini found two Claude runs that split a $1,600 order into four $400 payments under a $500 cap. The coffee template does not ship a velocity cap by default, and Pink scored its own runs.
Reality
Diogo Almeida left OpenAI two years ago convinced that human language is the wrong output for automation. The model his startup shipped this week returns probabilities, and one team testing it clocked classification 5 to 18 times faster.
Reality
Diogo Almeida's TypeSafe put Jev into early access on 15 September with $40m from DCVC, and it sells a calibrated confidence number on every answer as the thing that makes automation possible, with the evaluations behind that claim built in-house.
Reality
Coralogix CEO Ariel Assaraf says a misconfigured Gemini test agent entered three real systems before it stopped itself. His fix is a policy check between the agent and its tools, and the model has no power to override it.
Perspective Coverage
4 publishersReality
sdlc-playbooks runs a gate script before each of its seven coding-agent phases, and a failed precondition aborts the Claude Code command with exit code 2. The check binds only at that entry point, so direct script calls and other editing paths still need their own controls.
Reality
agent-tooltrust ran 206 live model calls instead of 2,490 and reports identical coverage, on the argument that the only thing a real LLM adds is proof that each framework adapter can surface all four verdicts.
Reality
The expensive part of software work has moved from writing code to proving it. An engineer building agent systems puts the deny outside the model where no prompt can reach it, and says plainly which defects his deterministic checks still miss.
Reality
Rulestack says only one of the five was the kind of error a string matcher could catch, and the first matcher it wrote cleared the draft anyway, because 76 turns up 1,343 times in the ledgers it searched.
Reality
A dev.to guide scores every action an AI legal agent takes by reversibility, blast radius and stakes, then attaches controls by grade: maker-checker on anything executed or filed, attorney review on all substantive work product.
Reality
The 18-hour-old cost-guard-mcp gets a real byte count out of BigQuery's dry-run planner and only a planner ceiling out of Snowflake's EXPLAIN, so its type system makes an agent see which of the two it got.
Reality
The gate between Claude Code's self-written audit memos and one developer's Obsidian wiki scans only the four files it intends to copy, tests them against two credential prefixes, and exits 2 on the first match.
Reality
Tenable's account of the control layer around Hexa AI lists what broke during development, including an agent that did not know who was asking, a model that denied tags it could not find, and queries that died past 5,000 assets.
Reality
One developer counted the automated checks in his own repositories and found that only about one in nine could prove it was capable of rejecting anything.
Reality
An engineer at Snappfood watched pull requests get bigger and faster after adopting AI tools. The part of review that catches bugs still works; the part that moved knowledge has no author to draw from.
Reality
A dev.to post makes a narrow, useful point: agent hooks can read intent, and cost is not a property of intent. The check belongs where the ledger already is.
Reality
Judge Walter Spader Jr. found that no Connecticut court uses AI on filings, then sanctioned the litigant anyway. A clerk caught the size-3 white text by its spacing, not by scanning.
Reality
PostHog asked Claude, Cursor and its own assistant the same revenue question and got three answers. Its remedy was a governed catalog of definitions stored as ordinary SQL tables.
Reality
The author of a now-public merge-gate core recounted his own suite: 22 tests in the evasion file, of which 17 are attacks and 5 exist to stop the gate being red all the time.
Reality
A PreToolUse hook on an AI coding agent returned "allow" for every file under a non-ASCII directory name, for weeks, with no log line and no exception.
Reality