Cybersecurity

19 articles tagged "Cybersecurity"

Anthropic's Claude Opus 4.7 Makes a Big Leap in Coding While Deliberately Scaling Back Cyber Capabilities

Anthropic has released Claude Opus 4.7, a new frontier model with significant coding benchmark improvements — paired with a rare deliberate reduction of the model's cybersecurity capabilities, signaling how the company is navigating dual-use AI risk at the frontier.

D.O.T.S AI NewsroomApr 17, 20264 min read

Startups

Gitar Emerges From Stealth With $9M to Use AI Agents to Secure AI-Generated Code

Gitar has launched from stealth with $9 million in funding, building AI agents designed to review and secure code that has itself been generated by AI — addressing a fast-growing security blind spot.

D.O.T.S AI NewsroomApr 16, 2026

Policy

Claude Mythos Can Autonomously Compromise Weakly Defended Enterprise Networks End-to-End, Anthropic Red Team Confirms

Internal red-team testing conducted before Mythos's restricted release has confirmed that Anthropic's most capable cybersecurity model can autonomously execute multi-stage network intrusion campaigns against enterprise environments with minimal human direction — a capability that marks a qualitative shift in AI-enabled offensive security.

D.O.T.S AI NewsroomApr 15, 2026

Policy

AI Agents Store Credentials in the Same Box as Untrusted Code. Two New Architectures Show How to Fix It.

A new VentureBeat analysis examines two emerging architectural patterns — from Anthropic and NVIDIA's NemoClaw — that attempt to isolate AI agent credentials from the untrusted code agents execute, addressing what security researchers describe as one of the most underappreciated risks in enterprise agentic deployments.

D.O.T.S AI NewsroomApr 13, 2026

Industry

Anthropic's Glasswing Initiative Confronts AI's Most Uncomfortable Security Paradox

Anthropic has launched Glasswing, a new security-focused AI initiative that attempts to harness the same capabilities that make AI dangerous for offensive purposes as a defensive tool — walking a careful line that reflects the dual-use dilemma at the heart of AI security.

D.O.T.S AI NewsroomApr 12, 2026

Policy

Anthropic Launches Mythos, a Powerful Cybersecurity AI Available Only to a Vetted Few

Anthropic has released Claude Mythos Preview, a specialized AI model designed for offensive and defensive cybersecurity work, with access restricted to a short list of approved organizations including Amazon, Apple, Microsoft, Broadcom, Cisco, and CrowdStrike. The launch signals a new category of AI deployment: frontier models too dangerous for general release but too valuable to leave undeployed.

D.O.T.S AI NewsroomApr 9, 2026

Policy

Anthropic's Claude Mythos Found Thousands of Zero-Days — So They're Not Releasing It

Anthropic has quietly restricted its most capable new model, Claude Mythos, after the system autonomously discovered thousands of critical vulnerabilities in major operating systems and browsers — including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw. The model is being deployed exclusively through Project Glasswing with 11 vetted security partners. It is the most concrete case yet of an AI lab withholding a model because of genuinely demonstrated risk.

D.O.T.S AI NewsroomApr 9, 2026

Policy

Hackers Are Redistributing the Leaked Claude Code Repository — With Bonus Malware Attached

Wired reports that threat actors are repackaging the leaked Claude Code source repository and uploading it to file-sharing platforms bundled with information-stealing malware. The pattern is a textbook social engineering play: developers curious about the leaked AI tool are downloading what looks like the genuine repository and executing malware in the process.

D.O.T.S AI NewsroomApr 8, 2026

Breaking

Anthropic's New 'Mythos' Model Found Security Vulnerabilities in Every Major OS and Web Browser — Now It's Powering Project Glasswing

Anthropic is deploying Claude Mythos Preview — a new general-purpose model it's not releasing publicly due to security concerns — through Project Glasswing, a defensive cybersecurity initiative with Nvidia, Google, AWS, Apple, and Microsoft as launch partners. Mythos reportedly found vulnerabilities in every major operating system and web browser during testing, making it too capable to release openly while too valuable to leave unused.

D.O.T.S AI NewsroomApr 8, 2026

Research

AI Offensive Cyber Capabilities Are Doubling Every Six Months, Safety Researchers Find

A new study from AI safety researchers shows that AI models' offensive cybersecurity capabilities are doubling approximately every 5.7 months. Current systems can now solve tasks that would take a skilled human security researcher around three hours — a capability threshold that didn't exist in commercial AI two years ago.

D.O.T.S AI NewsroomApr 7, 2026

Policy

AI Offensive Cyber Capabilities Are Doubling Every Six Months, Safety Researchers Warn

A new safety research report finds that AI models' ability to autonomously exploit security vulnerabilities has been doubling roughly every 5.7 months since 2024 — a rate that is outpacing the development of defensive tooling and policy frameworks.

D.O.T.S AI NewsroomApr 6, 2026

Policy

AI Offensive Cyber Capabilities Are Doubling Every Six Months, Safety Researchers Find

A study from Lyptus Research found that AI's ability to exploit security vulnerabilities has been compressing its doubling time from 9.8 months to 5.7 months since 2024 — with frontier models now able to complete tasks requiring three hours of expert human effort at a 50% success rate.

D.O.T.S AI NewsroomApr 6, 2026

Research

AI Offensive Cyber Capabilities Are Doubling Every Six Months, Safety Researchers Find

A new safety research study finds that AI models' ability to exploit security vulnerabilities has been doubling every 5.7 months since 2024 — with Anthropic's Opus 4.6 and OpenAI's GPT-5.3 Codex identified as the current capability frontier. The findings arrive as the AI safety community debates how to govern models that are now genuinely dangerous in the wrong hands.

D.O.T.S AI NewsroomApr 6, 2026

Research

AI Offensive Cyber Capabilities Are Doubling Every Six Months, Safety Study Finds

A new safety study finds that AI models' ability to exploit security vulnerabilities has been doubling every 5.7 months since 2024. Opus 4.6 and GPT-5.3 Codex can now autonomously solve cyberattack tasks that take human experts roughly three hours — and the curve is not flattening.

D.O.T.S AI NewsroomApr 5, 2026

Policy

Open Source Supply Chain Attack Hits AI Ecosystem: LiteLLM Compromise Leads to Mercor Data Breach

A cyberattack on AI hiring startup Mercor has been traced to a compromised version of LiteLLM, one of the most widely used open source AI infrastructure libraries. The incident is a sharp warning about the security posture of the rapidly growing ecosystem of AI tooling — where trust in open source packages is high and security scrutiny often isn't.

D.O.T.S AI NewsroomApr 2, 2026

Research

Google DeepMind Maps Six Ways Attackers Are Already Hijacking AI Agents

A new DeepMind study catalogs the first systematic threat model for autonomous AI agents, identifying six attack categories — from hidden HTML injection to multi-agent 'digital flash crash' attacks — with documented proof-of-concept exploits for every type.

D.O.T.S AI NewsroomApr 2, 2026

Policy

Mercor Confirms Data Breach Via LiteLLM Supply Chain Attack — A Warning Shot for AI Infrastructure Security

AI recruiting startup Mercor has confirmed a cyberattack tied to a compromise of the widely-used open-source LiteLLM gateway project, with a Lapsus$-affiliated extortion crew claiming responsibility — exposing a critical security gap in the AI startup ecosystem's reliance on shared open-source infrastructure.

D.O.T.S AI NewsroomApr 1, 2026

Industry

Okta's CEO Is Betting the Next Big Identity Problem Is AI Agents, Not Humans

Okta CEO Todd McKinnon is publicly framing AI agent identity as the company's defining next chapter — the infrastructure layer that must exist before agentic AI can be trusted in enterprise environments. The bet: every AI agent that touches enterprise systems needs credentials, permissions, and an audit trail, just like every human employee.

D.O.T.S AI NewsroomMar 31, 2026

Breaking

Leaked: Anthropic's 'Claude Mythos' Is the Most Powerful AI Model Ever Built

Internal documents accidentally exposed via a misconfigured content management system reveal Anthropic's next-generation model, codenamed 'Claude Mythos,' achieves dramatically higher scores in coding, reasoning, and cybersecurity than any previous AI system — including GPT-4o and Gemini Ultra.

D.O.T.S AI NewsroomMar 28, 2026