Security — Page 4

The `anthropickit` package was real malware; its unproven link to Anthropic points to a containment failure in agent evaluations.

The reported $1 billion transaction follows $126 million raised by two new platforms challenging employee-era identity stacks.

Attackers reached operational technology at more than 30 Minnesota utilities, briefly shutting controls at one treatment plant.

The August 14th rule covers new and existing accounts, while unpaid vulnerability disclosure programs remain open without verification.

A Claude model reached the internet from testing environments, turning controlled security exercises into unauthorized real-world access.

The model beats general-purpose rivals on Depthfirst's internal benchmark, whose construction and methodology remain unpublished.

Wiz says Microsoft fully remediated CosmosEscape, a Gremlin API vulnerability chain that reached a platform-wide signing secret.

Hakon Maloy says hidden prompts can alter reports and copy themselves into documents after two Microsoft mitigation attempts.

Bessemer led Onyx Security's Series B four months after Maxim Bar Kogan and Gil Elbaz took the Israeli AI security company out of stealth.

The new extortion actor published 15 victim claims in one burst, a pattern researchers say is more consistent with a bulk launch or fabrication.

Modal CTO Akshat Bubna says an exposed customer endpoint gave the escaped agent root access while Modal's platform stayed intact.

The agents escaped a cyber benchmark, exploited a zero-day and took 17,600 actions before Hugging Face contained the intrusion.

The July 28th disclosure says Claude Mythos Preview improved attacks on a NIST post-quantum signature candidate and a reduced-round AES variant, with no reported production impact.

Unit 8200 alumni Jonathan Jacobi and Gal Niv are using public interactions to test interfaces and collect training data for robot-agnostic AI.

The new model claims a 96% CyberGym score and half‑the‑cost operation, and debuts alongside the Perception agentic security system.

The reported 90.9% CyberGym score puts system design ahead of model choice, though Wiz has not published Atlas methodology or access details.

In a July 26th X thread, ex-Anthropic staffer Adi Baradwaj claims the company traded safety for revenue and notes that most black‑hat hackers use standard Claude Code subscriptions.

The X user rolled out a service that lets anyone access publicly exposed models such as Kimi 1T and Deepseek 765B.

After revealing a vulnerability that let anyone forge YC founder scores, partner Jared Friedman quickly fixed the flaw and extended a Summer 2026 Startup School invitation.

In a July 24th, 2026 X post, Musk pledged total transparency for the platform by releasing its code and subjecting it to external review.

The new model, now on all paid Claude plans, hits state‑of‑the‑art scores on coding, ARC‑AGI‑3 and alignment tests, according to Anthropic.

Cheyenne Ventures and AVP co-led the round as Abstract cited 380% ARR growth without disclosing its revenue base or dollar valuation.

Chaofan Shou says Kimi K3 found 19 flaws in 90 minutes; public code documents an authenticated exploit against Redis's latest release.

Clement Delangue's open-model bet became incident-response infrastructure after hosted APIs rejected the attack logs.