Thursday, 29 October 2026
Anti-bot fingerprinting scripts embed AI jailbreaks; Microsoft Edge falls to pure logic-bug chain; AI auditors find 2,000+ errors in 138 economics papers
Today's Lead
Engineeringnullpt.rs
Why Is This Fingerprinting Script Creating Bioweapons?
A security researcher reverse-engineers browser fingerprinting scripts by crawling roughly 10,000 websites and collecting their canvas renderings. One script from servicepipe.tech turned out to contain strings specifically designed to bypass AI safety filters. When a large language model reads the script for analysis, the embedded prompts redirect it into generating biosecurity content instead of a code review. The script itself is a standard anti-DDoS tool that traces to a Russian company — the AI-jailbreak strings appear to be a deliberate anti-analysis technique, making automated code auditing produce misleading output rather than a useful assessment.
Engineering
DEVCORE
Forging a Microsoft Edge Full Chain with Logic Bugs Only
Orange Tsai of DEVCORE built a full-chain exploit against Microsoft Edge for Pwn2Own Berlin 2026 using five logic bugs and two standard Windows features — no memory corruption and no AI assistance. The chain starts with a zero-click universal cross-site scripting vulnerability and ends with arbitrary code execution. The author argues that logic bugs are hard to discover but easy to chain once found, and suggests AI may become a meaningful aid for finding them in future security research.
GitHub Gist
Claude Code burns ~15–20% of your tokens on a 3-word status line
Claude Code sends the full transcript of each background subagent to a model on a recurring timer — roughly every 30 seconds — to generate a short status summary. The author measured this at around 15–20% of total token consumption. No official setting disables the behaviour. The author provides a Bun preload script that replaces the summary timer with an inert no-op, restoring the tokens. The workaround may break after a Claude Code update.
divagr.com
Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5
Divyansh Agrawal modified Qwen 3.5 4B so that a compiler sets model weights to embed a virtual 32-bit RISC-V processor inside the forward pass. When the model emits a tool call such as calc, put, or get, added Gated DeltaNet layers intercept it, execute the instruction on the virtual CPU, and write the synthesised response back into the token stream. The CPU uses GDN layers as exact, deterministic RAM. Long-integer arithmetic accuracy rises from 48.6% to 100%.
Alpine Linux
GitLab security incident: attempted exploitation of CVE-2026-97229
An attacker sent crafted CI configurations to Alpine Linux's self-hosted GitLab server on 20 and 21 October, targeting CVE-2026-97229, a critical bug in Ruby's regular-expression engine. The attack crashed GitLab worker processes eight times. Alpine found no evidence of code execution but treated the server as compromised anyway: they rebuilt it from scratch, reset all credentials, and replaced all secret keys. Self-hosted GitLab operators must update the underlying Ruby runtime separately — a GitLab upgrade alone does not close the vulnerability.
Humanities
JSTOR Daily
How Telepathy Helped Build the Russian Internet
Media archaeologist Natalia Konradova's book Hacking the Iron Curtain traces the origins of the Russian internet to mid-20th-century Soviet parapsychology research. The same community of scientists and citizen diplomats who studied telepathy as a plausible technology also organised the 1982 US-Soviet Space Bridge and, in the mid-1980s, invited American entrepreneur Joel Schatz into Soviet scientific institutes — enabling the first Usenet connection from Soviet territory. Konradova argues that the desire for unmediated mass communication drove both parapsychology and early networking, and that the Soviet internet was less a cause of the USSR's collapse than a symptom of the same political opening.
Aeon
The physics of getting fired
Zahaan Bharmal writes about being made redundant from a London consulting job in 2001 and the intellectual path that followed. He argues that physics trained him not just in equations but in a mode of reasoning — building simplified models, stress-testing assumptions, finding the hidden invariant. Applied to life, this meant understanding his firing not as a random blow but as an event in a chaotic system where small inputs cause large outputs. He advocates teaching physics as a way to reason about uncertainty rather than as a collection of formulas.