· ·

AI Security Newsletter No.91 (11-11-2026)

By Song Luo · AI security product leader

Welcome to this edition of the AI Security Newsletter. This week, frontier offensive models reached defenders under tight controls: OpenAI released GPT-6 Cyber to vetted Daybreak Red customers, and Cisco is adding Anthropic’s Claude Mythos to AI Defense red teaming. Open-source tools kept pace, led by 0.security’s open-source 0 harness and the sapper-v2 multi-agent pentest harness. We also cover the widening Fireworks AI breach, CrowdStrike’s AI kill switch, and new ways to govern what agents can reach.

Risks & Security

Fireworks AI Breach Widens to Customer Datasets and Model Weights

Fireworks AI has widened its October breach disclosure. It first reported that attackers accessed API keys and tokens in evaluation jobs, but now says attackers also reached account and billing information and, for a limited set of accounts, secrets, datasets and incomplete model weights. On October 24, they pivoted via a compromised third-party cloud provider into Fireworks development machines. Fireworks links the attack to a financially motivated “LLMjacking” group that sells stolen model access to illicit routing gateways.

References:

Technology & Tools

GPT-6 Cyber: OpenAI’s Model for Authorized Security Work

OpenAI released GPT-6 Cyber on November 5, a GPT-6 Astra variant post-trained to find vulnerabilities, develop exploits and validate patches. OpenAI says researchers used it to develop 948 Linux privilege escalation exploits and find 81 previously unknown high-severity Firefox bugs. In testing, it avoided unauthorized targets 99.93% of the time without deployment safety instructions and every time with them. Access is limited to approved Daybreak Red customers, with SSO, MFA and API IP allowlisting required.

References:

0.security Open-Sources the 0 AI Security Harness

Y Combinator-backed 0.security open-sourced 0, an AI security harness for code analysis, pentesting, vulnerability research, reverse engineering and fix generation. It runs as a local browser console, a CLI or an MCP server for coding agents, and works with your own model provider. The team says it has found more than 30 Linux kernel zero-days. Its best self-reported run solved 93 of 95 XBOW web CTF challenges, while a 2026 cold run solved 45 of 104.

References:

sapper-v2 Multi-Agent Pentest Harness Wins Tencent Qualifier

sapper-v2 is an open-source multi-agent pentest harness in which a commander agent dispatches scouts and soldiers that share a layered fact-intent graph. It won the 634-team online qualifier of Tencent Cloud’s third Intelligent Attack-Defense Challenge, running eight hours unattended and scoring 80.18 out of 100. It matched the runner-up’s 85 flags 25 seconds faster, on 1 billion tokens versus 4.5 billion, by using stronger models for planning and cheaper ones for bulk work.

References:

Enterprise Managed Auth Connects Claude to MCP Servers Without Consent Screens

Claude’s Enterprise Managed Auth, on Team and Enterprise plans, lets organizations connect users to MCP servers without a per-user OAuth consent screen. Claude presents the server’s authorization server with a signed identity assertion, issued by the customer’s identity provider or by Claude, and receives an access token. Built on the MCP Enterprise-Managed Authorization extension, it lets admins authorize a connector once while users inherit access through IdP groups, and short token lifetimes make revoked access expire quickly.

References:

Mozilla AI Pitches Otari as a Governance Gateway for Agents

Mozilla AI is positioning Otari, its open-source AI gateway, as a governance layer for apps and agents. Per workspace, teams set allowed models, turn web search and code execution on or off, and require guardrail checks, with flagged requests blocked before reaching a provider. Budgets apply to users and application keys. Agent Guardrails can refuse specific tool calls from coding agents, and session tracing shows what ran and what it cost without storing prompts.

References:

Business & Products

Cisco AI Defense Adds Anthropic’s Claude Mythos for Red Teaming

At Cisco Live APJC in Melbourne, Cisco said it is adding Anthropic’s Claude Mythos to the adaptive red teaming framework in AI Defense. Mythos plans multi-step attacks tailored to each customer’s AI agents and applications. Customers get no open-ended prompt access: tests run only against registered, authorized targets, each engagement is tied to a named operator, and every action is logged. Access starts with select enterprises, with no general release date yet.

References:

CrowdStrike Adds an AI Kill Switch to Falcon Guardian

At Fal.Con Europe, CrowdStrike added an AI Kill Switch to Falcon Guardian. Security teams can use it to direct connected AI gateways, either the now generally available Falcon AI Gateway or a third-party gateway, to block the models an agent depends on. The first version cuts off model access rather than stopping the agent’s process; endpoint enforcement and wider containment are planned. Guardian also adds agent-level hooks for tools such as Claude Code, Codex and Cursor.

References:

Leave a Reply

Discover more from Mindful Machines

Subscribe now to keep reading and get access to the full archive.

Continue reading