A nova campanha da Vyntra já está ativa: Operation Erebus:
https://t.co/C64bpTW2e6
Ela expande o modelo de pesquisa da anterior Acheron para uma abordagem de longo prazo focada especificamente em superfícies low-level, runtime/native interfaces, trust transitions e desenvolvimento de attack chains.
A operação roda em 3 frentes independentes, cada uma com um objetivo principal: confirmar 1 zero-day crítico de RCE.
Até agora, a campanha já produziu 2 findings High no front "Nullcrown", incluindo uma falha de confidencialidade/integridade em respostas protegidas e uma condição de resource exhaustion não autenticada com impacto service-wide.
Os detalhes técnicos continuam restritos durante a triagem e coordinated disclosure.
New technical writeup published on CVE-2026-75604, a critical Next.js vulnerability affecting Windows-hosted applications.
The issue combines a cache path traversal with access to internal server metadata, which can ultimately lead to unauthenticated remote code execution in affected configurations.
Full technical analysis:
https://t.co/0XbhIIncWU
We’ve launched the Vyntra Operations blog to publicly track the progress, status and outcomes of our active security research campaigns
The current highlight is Operation "Acheron", our first campaign applying the full set of lessons from advanced vulnerability research and hybrid AI-assisted workflows, with a specific focus on high-impact flaws in widely used products.
So far, Acheron has produced 5 findings, including 1 critical server compromise at an administrative level.
The goal of the Operations blog is to make this process visible as it evolves, including active targets, campaign status, validated findings and future disclosures.
https://t.co/CPD3jAZsMT
The core objective of this initiative is to find critical vulnerabilities by combining refined human heuristics with precisely directed autonomous research.
Over the past months, we have been improving a methodology designed to push agents beyond shallow pattern matching and toward complex, previously unexplored vulnerability chains. Instead of asking a model to broadly “find bugs,” the workflow gives it a progressively refined understanding of the target, its architecture, its trust boundaries, its historical weaknesses and the hypotheses that are still worth pursuing.
A central part of this process is the research corpus.
The corpus is the persistent body of knowledge accumulated throughout the campaign. It records what has already been analyzed, which hypotheses were tested, why certain paths were discarded, which components remain promising, what technical constraints were discovered and how previous findings changed our understanding of the target.
This means each new research cycle starts from a stronger baseline instead of repeating earlier work.
Our Proteus framework is used to persist and share this knowledge across the workflow, allowing autonomous agents and human researchers to operate from the same evolving context. Explored paths are not unnecessarily repeated, negative results remain useful, and observations from one stage can directly improve the steering of the next.
Human work remains active throughout the process.
Parallel intelligence gathering helps calibrate the campaign through the target’s CVE history, prior architectural failures, public patches and related products. This material is then converted into more specific playbooks and heuristics, helping identify where complex chains are more likely to exist and which assumptions deserve deeper validation.
We refer to this as directed human-in-the-loop research.
The autonomous side provides speed, execution capacity and the ability to explore large codebases, build labs and validate multiple hypotheses. The human side defines the direction, evaluates impact, recognizes promising inconsistencies and continuously adjusts the research strategy.
The result is a workflow designed to avoid both extremes: purely manual research that cannot scale, and expensive agent swarms producing large volumes of duplicated or superficial output.
This initiative is another step toward building a highly effective research system that does not depend on spending thousands of dollars on tokens. The goal is not to maximize the number of agents running at once, but to maximize the quality of every research cycle by combining persistent knowledge, precise steering, human judgment and autonomous execution.
New research published: CVE-2026-64648, discovered by @sybr1d_ in Next.js:
https://t.co/G8hfwLp7Uw
The issue affected fetch(Request, init) caching: Next.js could generate the cache key from a different request than the one actually sent upstream, allowing a cached POST response associated with one user to be returned to another.
Estamos lançando oficialmente os serviços de pentest da Vyntra.
https://t.co/T7iytoyglr
Além dos artigos e writeups sobre vulnerabilidades que encontramos, agora devs e times menores também podem contratar esse mesmo nível de pesquisa especializada para avaliar a segurança dos próprios produtos.
Nosso foco inicial está em aplicações web e código-fonte, com uma precificação transparente, padronizada e acessível também para produtos menores, projetos independentes e aplicações vibe coded.
A ideia é simplificar o acesso a uma análise de segurança mais profunda, com preços claros e um processo de contratação direto e previsível.
A Vyntra segue com o mesmo princípio desde o início: tornar conhecimento, ferramentas e pesquisa em segurança mais acessíveis e públicos.
Uma nova pesquisa da Vyntra Research demonstra como um post no X pode ser transformado em uma cadeia completa de prompt injection e exfiltração de conversas privadas no Grok:
PT: https://t.co/PjL00ZCjiT
EN: https://t.co/cKfzRUVrKu
O ataque combina Unicode invisível, um primeiro estágio embutido no conteúdo público e um segundo payload carregado externamente, permitindo ocultar a lógica maliciosa e alterar instruções de forma progressiva.
invisible Unicode payload
↓
first-stage prompt injection
↓
external second-stage instructions
↓
access to private conversation context
↓
data encoding
↓
external fetch used as an exfiltration channel
A pesquisa mostra como capacidades legítimas do assistente podem ser encadeadas em um fluxo completo de distribuição, controle e exfiltração.
Publicamos um novo artigo na Vyntra Research sobre uma vulnerabilidade encontrada por nosso pesquisador, @sybr1d_, no MCP Client do Vercel AI SDK:
PT:
https://t.co/iOaUu1IKgk
EN:
https://t.co/4XBICmsu5J
A falha permitia que um MCP server malicioso contornasse a allowlist de tools definida pela aplicação e expusesse ao modelo uma tool que não havia sido aprovada.
O problema estava em uma checagem feita com o operador in do JavaScript, que também considera propriedades herdadas pela prototype chain. Com isso, uma tool chamada constructor podia atravessar a validação, ser enviada ao modelo e executada via tools/call.
No artigo, detalhamos o fluxo vulnerável, o impacto, a causa raiz e por que own-property checks são importantes quando objetos JavaScript participam de decisões de autorização.
Hi, we’re Vyntra Security.
We created Vyntra to make research-grade offensive security more practical and accessible for modern software teams, without reducing pentesting to automated scans or generic checklists.
Our research division, Vyntra Research, is already live with three articles, including technical writeups for CVE-2026-44575 and CVE-2026-40073, both written by our researcher @sybr1d_.
Vyntra Research will publish vulnerability research, CVE writeups, exploitation details, tooling, and the methods behind our work.
More is coming soon, including an article on another vulnerability found in the AI SDK and a deeper look at the hybrid human and AI research process we use to find and validate real security issues.
Read the first publications:
https://t.co/qQgKL2IANy