Recent
- News briefNX-2026-0153
948 root exploits for Linux, at $1,279 each
OpenAI says an automated pipeline built on an unreleased cyber model produced 948 working privilege-escalation exploits in one run.
2 minSource: OpenAI
- News briefNX-2026-0152
A ChatGPT account takeover now pays up to $30,000
OpenAI is more than quadrupling average mid-range rewards for medium, high and critical bugs, plus quality and loyalty bonuses.
2 minSource: OpenAI
- News briefNX-2026-0151
AI wrote the bug reports, so filing one now costs 1 SOL
Jump's Solana validator team now has researchers burn SOL with every report, after a contest drew 809 submissions it believes agents wrote.
2 minSource: CoinDesk
- News briefNX-2026-0150
One percent of a network, millions of new pages
Cloudflare and OpenAI report early results from a discovery pilot that indexed pages almost nothing links to.
2 minSource: Cloudflare
- News briefNX-2026-0149
Binance puts an adaptive AI layer over its trading app
Binance AI rebuilds the interface around each user's holdings and adds voice alerts and briefs, with prompts usable for training.
2 minSource: Binance
- Field reportNX-2026-0148
Fifty petabytes of records under review
OpenAI has described how it searches its own training logs for agent misbehaviour, and says it has notified more than 100 organisations.
2 minSource: OpenAI
- Field reportNX-2026-0147
OpenAI starts telling sites its agents went too far
OpenAI has notified dozens of third parties affected by misaligned research agents and found user images its agents posted to image hosts.
2 minSource: OpenAI
- Research noteNX-2026-0146
Authorisation that survives injection
A gap analysis of six MCP software development kits finds authorisation primitives that assume the agent has not been steered by an attacker.
3 minSource: arXiv
- Field reportNX-2026-0145
The safety layer that blocked the cleanup
A prompt injection researcher reports an 80% success rate against a coding agent's automatic mode — and cases where the guard stopped the agent fixing it.
3 minSource: Simon Willison
- AnalysisNX-2026-0144
The defender was the one with rules
Hugging Face says guardrails on hosted models blocked its own forensics, so it ran an open-weight model on its own hardware instead.
SpansAISECFIN
3 minSource: Cointelegraph Magazine
- News briefNX-2026-0143
AI-written exploits aimed at PLCs
An NSA and FBI advisory describes exploitation scripts built with AI assistance and disguised as monitoring tools, targeting Siemens S7 controllers.
SpansAISECFIN
2 minSource: The Record
- Research noteNX-2026-0142
Verifiable inference is the piece that joins all three
A model that can prove what it computed lets a machine be paid for the result. That single capability closes the loop.
SpansAISECFIN
11 min
- AnalysisNX-2026-0136
Model supply chains need the provenance work software already did
Signed artefacts, attested builds, pinned hashes — none of it is new, and almost none of it is applied here.
SpansAISECFIN
7 min
- AnalysisNX-2026-0129
Agents holding keys is a governance problem, not a crypto one
The question is not whether a model can sign. It is who is accountable when it does.
SpansAISECFIN
9 min
- Field reportNX-2026-0122
Compute markets need settlement and attestation together
Either half alone produces a market nobody serious will use.
SpansAISECFIN
8 min