We Built An Agent That Broke All of Our Security Rules. Here's What We Learned.

thehackernews.com Nov 11, 2026

We Built An Agent That Broke All of Our Security Rules. Here’s What We Learned.

Curated by Shubham · TechCapsules · Nov 11, 2026

AI-summarised brief · reviewed before publication

Token created an AI browser agent named Rehearsal that intentionally violated its own security guidelines—no distinct identity, unrestricted access, and no traceability—to expose weaknesses in agent protection. The agent operated within users’ authenticated sessions, performing tasks like opening pages and filling forms, making its actions indistinguishable from human clicks. By analyzing this failure, Token identified gaps in current practices and highlighted the need for separate agent lanes, identity tagging, and scoped permissions. The company outlined how emerging browser features and standards could address these issues.

💡 Why It Matters

· The experiment shows that without dedicated agent identities and scopes, automated tools can bypass human oversight, turning legitimate workflows into potential attack vectors.
· It underscores the urgency for immediate, practical safeguards while industry standards mature.

AI Agents Capsules

Read full article on thehackernews.com

Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than…

Microsoft researchers demonstrated that an AI model can be backdoored by editing fewer than ten rarely used neurons, without retraining. The technique, named Lazarus, allows a trigger string to cause the model to output an attacker’s command, which an agent can then execute with its own permissions. Tests on three open‑weight models—Nemotron‑Terminal‑8B, Qwen3.6‑27B, and Gemma‑4‑31B‑IT—showed success rates up to 99.5% in triggered sessions while remaining functional otherwise. Microsoft has not observed real‑world use and released defensive guidance.

💡 The method reveals that minimal, targeted weight changes can covertly compromise AI agents, posing a low‑effort threat to any system that can alter model checkpoints before deployment.

9to5mac.com · Nov 8

Apple aiming to make Siri ‘more like an agent’ as soon as…

Apple released a personalized Siri with iOS 27, fulfilling its 2024 roadmap but still lacking proactive task execution. Bloomberg’s Mark Gurman reports that Apple plans to introduce an agent‑style Siri as early as next year, enabling the assistant to track goals, decide next steps and act without user prompts. The push follows similar moves by Meta, SpaceXSI and emerging agents such as Muse and Grok Bot, which run their own virtual machines to handle background tasks. Apple’s extensive App Intents framework could reduce user reliance on manual computer interaction, making complex workflows cheaper. To accelerate development, Apple is extending its coding phase from four to six weeks in the iOS 28 cycle, giving engineers more time for exploration before a shortened two‑week integration period.

💡 A proactive Siri could redefine everyday computing by automating multi‑step tasks, giving Apple a foothold in the fast‑growing agentic‑AI market.

techcrunch.com · Nov 5

Axiom Math wants to keep AI slop out of mathematics

Axiom Math, a $1.6 billion startup, announced on Wednesday that it has formally verified the Stacks Project, a 7,654‑page open‑source compendium of algebra, geometry and number theory. Using its own AI agents, the company translated the library into roughly 16 million lines of Lean code, preserving every logical connection across the massive hierarchy of theorems. The effort required a loosely structured, parallel workflow that let agents self‑assign tasks and develop a “taste” for mathematical rigor. CEO Carina Hong, a former Stanford Ph.D. dropout, said the verification fills a long‑standing gap in formalized mathematics, especially in algebraic geometry, and creates reusable building blocks for future AI‑driven proofs and software verification. Axiom raised a $200 million Series A in March to fund the project.

💡 By turning the Stacks Project into machine‑checked code, Axiom gives AI systems a reliable mathematical foundation, enabling them to generate provably correct software and scientific results.

quantumcomputingreport.com · Nov 4

IBM Quantum and QuICS Deploy AI Agents to Discover Over 500 High-Efficiency…

Researchers from IBM Quantum, the Joint Center for Quantum Information and Computer Science (QuICS), and MIT introduced a neuro‑symbolic search framework that couples large language model agents with deterministic integer‑programming solvers to explore magic‑state distillation protocols. Their preprint, “Symmetry and AI‑assisted discovery of magic‑state factories,” reports 699 factory classes, 564 of which are previously unknown, achieving record‑low physical resource footprints for non‑Clifford state production. By proving that code distance ≥ 3 corresponds to selecting nonzero, pairwise‑distinct syndromes in a binary‑matrix representation, the team expanded the searchable space beyond the historic limit of n ≤ 54. The automated multi‑agent loop iteratively proposes symmetry orbits, constructs circuit matrices, verifies gate identities and distances, and redirects searches toward promising geometries. Verified circuits and an interactive catalog have been released openly for community use.

💡 The breakthrough dramatically reduces the qubit overhead traditionally required for fault‑tolerant quantum computation, bringing practical large‑scale algorithms closer to reality.

commandline.microsoft.com · Nov 2

Nine engineers, 1,211 code updates in a day: Lessons from building Copilot…

In January, a nine‑engineer Microsoft team launched Copilot Cowork, an AI‑driven execution layer that automates multi‑step workflows for knowledge workers using Work IQ, Microsoft 365 security, and a multi‑model architecture. Within four and a half months the project progressed from an empty repository to general availability, and on October 21 the team merged 1,211 pull requests in a single day. Most code was generated by AI agents, while engineers defined specifications, set architectural and security boundaries, and reviewed high‑risk changes. The team instituted six practices—writing markdown specs before code, confining narrow agents to isolated containers, testing automated checks, preserving learned knowledge, structuring the codebase for parallelism, and limiting automated loops—to maintain reliability, security, and cost. The final milestone was migrating the live runtime from Python to Rust while the service remained operational.

💡 Demonstrating that AI‑generated code can be safely scaled in production reshapes how large enterprises might accelerate feature delivery without sacrificing governance.

techcrunch.com · Oct 27

How OpenAI turned its coding tools into dots, an always-on personal assistant

OpenAI unveiled “dots,” an always‑on AI personal assistant, at its September developer day, positioning it as the next evolution of its Codex coding tool. Alexander Embiricos, head of the enterprise product team, explained that dots reframes the underlying large‑language model as a single, continuous conversation rather than discrete prompts, aiming for intuitive, proactive assistance. In testing, the assistant flagged calendar conflicts and even detected a drone‑delivery outage before the user noticed, though timing remains imperfect. Unlike Meta’s Muse, which builds detailed social profiles for advertising, dots conducts “proactive research” by periodically reviewing user data to suggest actions, raising similar privacy questions. OpenAI hopes the persistent agent will drive higher token usage while delivering a more helpful, background‑running experience for users.

💡 Embedding a continuously listening AI turns a chat model into a data‑rich, revenue‑driving service, forcing the industry to balance convenience with user privacy.