Skip to content

Instantly share code, notes, and snippets.

@tcz
tcz / _results.md
Created October 4, 2026 03:57
Social-psychology biases in LLM oversight monitors (bystander, in-group favoritism, conformity) - ControlArena PoC

Social-psychology monitor PoC

Method

  • Settings: Apps (111 tasks, stored backdoors), BashBench (257 tasks, agent trajectories generated by abliterated GLM-5.3, 95% attack success, benign payload).
  • Monitors: claude-haiku-4-5, gpt-4.1-mini, gpt-4.1. ControlArena default prompt + one sentence per condition.
  • Threshold fixed at 5% FPR on each setting's control condition, applied to all conditions.
  • Metrics: catch rate on attacks, FPR on clean, mean suspicion, AUROC. 95% paired bootstrap CIs over tasks. 3 epochs (Apps), 2 (BashBench).

1. Bystander

Scanned 12366650 URLs. 416 raw payload matches, 416 after dropping security/doc sites.
HOST HITS GEN PATTERNS
------------------------------------------------------------------------------
petrichor.circlebit.ae 26 26 sqli_sleep
abudhabiculture.ae 16 6 sqli_sleep
minebox.africa 6 6 sqli_sleep
html.am 5 5 traversal
slc.dubai.gov.ae 1 1 traversal
@tcz
tcz / ai_village_mentions_2026.py
Created September 30, 2026 13:16
AI Village Mentions 2026
"""
What predicts how often an AI Village agent gets mentioned by other agents? (2026 only)
For each month of 2026, agents with >= 50 chat messages are ranked within the month by
chattiness (own message count), time in the village, and capability (Epoch Capabilities Index).
Output: median "attention index" per within-month percentile band, where attention index =
agent's share of that month's agent-to-agent mentions x (N - 1), so 1 = fair share.
rho = mean within-month Spearman correlation; p = permutation test (values shuffled within month).
Data:
Processing 240 countries...
[1/240] Processing Afghanistan (AF)...
Created: 1216_poster, unframed, standard - Cost: 875 GBP, Price: 0 (7-16 days)
Product in AF not available for 1216_poster, unframed, priority: Product not available for order: 2c7fea09-e90c-4a1d-9de9-2c0c5c4d3137, {"outcome":"NotAvailable","issues":null,"quotes":[],"traceParent":"00-696b7e9710bc9f8baaec3d2f70d0793c-a97bc44321760599-00"}, {"shippingMethod":"Express","destinationCountryCode":"AF","items":[{"sku":"GLOBAL-FAP-12x16","copies":1,"attributes":{},"assets":[{"printArea":"default"}]}]}
Created: 1216_poster, framed, standard - Cost: 6495 GBP, Price: 9000 (1-3 days)
Skipping 1216_poster, framed, priority - No previous shipping option found
Created: 1824_poster, unframed, standard - Cost: 963 GBP, Price: 0 (7-16 days)
Product in AF not available for 1824_poster, unframed, priority: Product not available for order: 9eeeffcb-11b0-4328-baba-1060f4cff0ec, {"outcome":"NotAvailable","issues":null,"quotes":[],"traceParent":"00-f94eb7
<?php
class DrinkingGame
{
const TAIL = 'tail';
const DRUNKEST = 'zoltan';
public function __construct(array $names, Coin $coin = null) {
$this->names = array_fill_keys($names, 0);
$this->coin = $coin;
@tcz
tcz / MainActivity.java
Created April 16, 2015 16:55
Android bug OpenGLRenderer﹕ Layer exceeds max. dimensions supported by the GPU
package com.example.gpulayerbug;
import android.app.Activity;
import android.os.Bundle;
import android.transition.Fade;
import android.transition.Transition;
import android.transition.TransitionManager;
import android.view.View;
import android.view.ViewGroup;
import android.widget.Button;
@tcz
tcz / designer.html
Created June 3, 2014 15:47
designer
<link rel="import" href="../core-scaffold/core-scaffold.html">
<link rel="import" href="../core-header-panel/core-header-panel.html">
<link rel="import" href="../core-menu/core-menu.html">
<link rel="import" href="../core-item/core-item.html">
<link rel="import" href="../core-icon-button/core-icon-button.html">
<link rel="import" href="../core-toolbar/core-toolbar.html">
<link rel="import" href="../core-menu/core-submenu.html">
<link rel="import" href="../google-map/google-map.html">
<polymer-element name="my-element">
@tcz
tcz / guessgame_loop.py
Last active August 29, 2015 14:01
Benedek python practice
import random
import sys
enter = "Please Press Enter To Continue..."
print "Hello! Welcome to a Guessing game!"
print "Please Guess A number between 1 - 100!"
computernum = random.randint(1, 100)
computernum = str(computernum)
guesses = 3
<?php
use Symfony\Component\HttpFoundation\Session\Storage\NativeSessionStorage;
/**
* Session sotrage that avoids using _sf2_attributes subkey
* in the $_SESSION superglobal but instead it uses
* the root variable.
*/
class LegacySessionStorage extends NativeSessionStorage
@tcz
tcz / nginxrps.php
Created February 24, 2014 13:51
Check request per seconds on Nginx with PHP
<?php
while (true)
{
$query_start = microtime(true);
$curl = curl_init();
curl_setopt($curl, CURLOPT_URL, "http://127.0.0.1/nginx_status");
curl_setopt($curl, CURLOPT_RETURNTRANSFER, true);
curl_setopt($curl, CURLOPT_HTTPHEADER, array(