Friday, October 16, 2026

Back to today

From X

@StableBread

Fajasy on X

I listened to the CEO of Nubeva Technologies $NBVAF (TSXV: $NBVA), a ransomware recovery company, explain why AI data labeling companies came to his team. Data labeling companies (firms that turn experts’ knowledge into training data for AI labs) have asked Nubeva whether it could help them label what each malware binary does, like whether it’s secretly encrypting files. Malware often arrives as a binary, a compiled program in machine code, and many are obfuscated, or scrambled to hide their purpose. CEO Randy Chou said in an October 5, 2026 interview on Nubeva’s YouTube channel that Nubeva has spent the last 7-8 years learning to spot homemade encryption, meaning encryption that ransomware authors write themselves instead of using a standard method. Homemade encryption is harder to catch, since it doesn’t match any of the standard methods security tools know. Here’s how Chou described what frontier models (the most advanced models from labs like OpenAI and Anthropic) can figure out today: “Frontier labs out there with their cybersecurity fully enabled, when they do full automation, they can approximately figure out 30% of the malware binaries. That’s the published numbers that are out today.” That claim is in line with GPT-5.6-sol’s 30.5% score on SRE-Bench, a Columbia University-led test of AI models on 262 binaries, but OpenAI’s GPT-6 Astra has since reached 56.9%. Chou said that for the other ~70%, a model might flag a binary as weird but can’t say what it does. He mentioned that some malware arrives compressed and obfuscated, then checks whether someone opened five apps, like a browser and Microsoft Office, in the past 72 hours, a known sandbox-evasion technique. If not, the malware assumes it’s inside a sandbox (an isolated computer analysts use to test suspicious programs) and does nothing. Today, Chou said, a small group of specialist cybersecurity labs takes malware apart by hand, and analyzing a new ransomware family typically takes weeks or months. He added that Nubeva has tested binaries written by its own engineers that frontier models couldn’t explain. That’s why Nubeva is now training its own model, which has no internet access, with “hundreds and thousands of AI agents.” The training only just started, but Chou said the timeline Nubeva is working to is “days and weeks,” not the 6-12 months he’d have quoted before AI agents. “I’m putting all my focus on this now,” he said.

2026.42: Home At Last

Stratechery’s weekly bundle covers Apple HomeView and SpaceX’s cellular ambitions.

Is The Market Turning The Page?

CoreWeave’s debt and equity price action is framed as a wake-up call for investors amid a broader market discussion.

PesoRama: Research Flash

PesoRama will reach 54 JOi Dollar Plus stores by November after raising C$28.4 million, funding more than 50 additional Mexican store…

Modine Deep Dive

Modine, soon Modexus Solutions, has a $4 billion hyperscaler cooling contract but must ramp factory output and restore margins as data…

TY · long

Tri-continental CORP

Tri-Continental Corp is presented as a long investment-grade fixed-income idea for portfolios with limited liquidity and diversification.

AppLovin Deep Dive

AppLovin rebuilt its ad model with AXON 2.0, lifting LTM revenue to about $7 billion and margins to 77% as Unity competition threatens…

Constellation: Stars Aligning?

Constellation Brands trades near $127 at roughly 11x next-twelve-month EV/EBIT, with recent business results offsetting concerns over…

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 107

X accounts we follow

Every account on this site, hand-picked.

88 accounts

Investor letters we follow

189 firms

Sectors

Ordered by how often our editor reaches for them.