Sunday, October 25, 2026

Back to today

From X

@StableBread

Fajasy on X

If the AI model Nubeva Technologies $NBVAF (TSXV: $NBVA) is training to read malware works, it could also take over the one step in its ransomware recovery software that engineers still do by hand. Nubeva’s Ransomware Reversal software, or NuRR, copies the encryption keys during an attack, but someone still has to figure out how ransomware uses them before Nubeva can build a decryptor (a program that reverses the encryption). AI agents have run NuRR’s routine support (bug fixes, support tickets, and product maintenance) for some time, CEO Randy Chou said, but Nubeva’s engineers still handle each live ransomware case by hand. Nubeva’s first granted patent finds the encryption method by running the ransomware in a sandbox (an isolated test computer) and matching the result to a library of standard methods. But the sandbox method breaks down in two cases Chou described, since homemade encryption (encryption ransomware authors write themselves) won’t match any standard method, and malware that detects the sandbox does nothing. A model that reads the malware’s compiled code could get around both problems, since (1) homemade encryption still shows up in the code as steps a model can follow, and (2) the malware’s checks for a real computer only run when the malware runs, so reading the code never triggers the checks. With NuRR supplying the keys and the model working out the encryption method, Nubeva would have both inputs a decryptor needs without an engineer taking the ransomware apart by hand. Frontier models, the most advanced AI models from labs like OpenAI and Anthropic, can already work out what some binaries (compiled programs) do just by reading their code. Anthropic said in April 2026 that its Claude Mythos Preview model can take “a closed-source, stripped binary,” meaning one with its function and variable names removed, and reconstruct “(plausible) source code for what it does.” Obfuscation (scrambling code to hide its purpose) is harder for a model to get past, since compressed malware keeps its real code hidden until it runs. Chou blamed obfuscation for the ~70% of malware binaries frontier models can’t figure out, so Nubeva’s model still has to prove it can do better. Chou didn’t connect the model to NuRR in his October 5, 2026 interview, but working out the encryption method is the same step Nubeva’s FY2026 annual filings say its AI decryptor builder will automate. Therefore, if the model works on live ransomware, NuRR could take on more cases at once, easing the capacity problem that cost Nubeva its biggest customer in 2023 (risk advisory firm Kroll).

Unicycive (UNCY):

Unicycive’s oxylanthanum carbonate NDA earned a March 25, 2027 PDUFA date after two manufacturing-related rejections, with no FDA clinical…

Moon River Moly: Research Note

Moon River Moly’s updated Davidson PEA lifts after-tax NPV to $1.24 billion as flotation raises tungsten output 350% and cuts molybdenum…

GLOO, part two

Gloo faces related-party deals, insider-heavy IPO ownership, acquisition losses and $84.5 million of CEO-backed puts as it burns cash and…

Carecloud News

CareCloud may attract a private-equity buyer after its Q3 call, with acquiring enough shares the main obstacle to a take-private deal.

LXS.DE: a 2028 sale it can force

Lanxess can force Advent to buy at least half its Envalior stake in April 2028, potentially raising about €745 million to reduce net debt.

EVC · long

Entravision Communications

Entravision Communications, up fourfold year to date, is pitched as an undervalued growth business emerging from a troubled parent…

Build-A-Bear: A Narrower Path

Build-A-Bear's push into prestuffed toys and broader retail distribution tests brand stretchability, but recent results warrant caution.

Twenty Years in the Dark

Tower Properties, $TPRP, went dark after a 2006 squeeze-out cut holders below 300, leaving a Kemper-controlled apartment and office…

TSOH Weekly Roundup (10/16/2026)

Bumble agreed to a $1.3 billion take-private deal, while Netflix cut about 1,000 jobs and Meta discussed its Muse superintelligence models.

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 109

X accounts we follow

Every account on this site, hand-picked.

91 accounts

Investor letters we follow

190 firms

Sectors

Ordered by how often our editor reaches for them.