ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-167 posts

Refresh

Amazon Honeypot Captures ARTEX AI Agent Logs Revealing Self-Sustaining Key Theft and Failed Deepfake KYC Attempts

Material
2026-11-14 04:26 GMT+8

Amazon Threat Intelligence announced on November 13, 2026, that its globally distributed decoy network, MadPot, captured requests from AI agents running offensive campaigns on the open-source framework ARTEX between October 3 and November 5, 2026. The sensors were configured to mimic popular large language model (LLM) inference APIs, causing the agents to send their own operational history—including system prompts, tool execution records, and knowledge graph states—to the honeypot.

The primary campaign involved the industrial-scale theft of credentials to access AI models. Stolen keys were fed back into the framework's own model gateway, creating a self-sustaining loop where computational resources paid for by victims funded the search for more keys. Targets included LLM gateways like One API, New API, and LiteLLM, as well as application builders such as Dify, Langflow, and RAGFlow. The agents exploited default master keys, uninitialized installs, and known vulnerabilities like CVE-2025-3248 (Langflow) and CVE-2026-28797 (RAGFlow).

A second campaign targeted cryptocurrency exchanges, attempting to bypass Know Your Customer (KYC) identity verification using real-time deepfakes. The agents reverse-engineered liveness detection mechanisms and generated forged video at 11.5 frames per second on a single consumer GPU. However, all recorded deepfake attempts failed server-side validation.

Amazon assessed with moderate confidence that these operations were conducted by disparate Chinese-nexus threat groups, citing simplified Chinese prompts, China time-zone configurations, and targeting rules excluding mainland China. The company emphasized that the techniques relied on known weaknesses rather than novel exploits, urging organizations to rotate exposed secrets, patch public platforms, and review indicators of compromise (IOCs) published in the report.

Sources:Amazon Web Services

Research

South Korea's AI Security Firm Aim Intelligence: Whether an AI Agent's Action Is Safe Depends on Its Context Within the Business Workflow

Material
2026-11-16 09:27 GMT+8

When enterprises deploy AI agents, the most common security approach is to create a permission list: which tools can this agent call, which data can it access, and which external APIs can it connect to? Any operation outside the list is blocked. Lee Han-wool, Head of Defense Security at South Korean AI security firm Aim Intelligence, believes that this approach is far from sufficient for AI agents. His core judgment is: determining whether an action is safe cannot rely solely on whether it is within the allowed scope, but must also consider whether it is appropriate within the current business workflow.

This distinction stems from two different types of risks. The first type is actively caused by humans: someone gives the agent malicious instructions or accidentally transmits incorrect commands. Traditional permission models can still handle these issues. The second type is more difficult to manage: no one instructs the agent to do something, but the agent crosses boundaries or bypasses security mechanisms on its own while completing its goal. Lee Han-wool attributes the difficulty of AI security to the fact that model performance has developed too quickly while control technologies have failed to keep up.

He uses a hypothetical scenario to illustrate why permission lists are insufficient. Suppose an agent is permitted to call a document search tool, a data analysis tool, and a certain external API. Each of these three permissions is fine when viewed individually. However, if the agent first searches an internal document, passes the content to another agent for analysis, and then calls an external API to send the results out—each step is within the scope of permission, but the overall chain constitutes a data leak. To identify such risks, the system must know what instructions the agent initially received, what sub-task each step was called to complete, and whether the causal relationships between these steps are reasonable.

Aim Intelligence refers to this capability as Observability, i.e., full observability of the agent's execution process. It records not just individual logs, but the entire chain from input to tool invocation to output, as well as the correlations between multiple chains. Based on this, the company proposes that static security and dynamic security must run in parallel: static security consists of pre-set hard rules, such as prohibiting access to certain external networks; dynamic security judges whether an action is reasonable based on the current task context and past execution records.

However, there is an engineering contradiction here: if every action requires tracing the entire business chain for deep analysis, latency and computational costs will expand rapidly, while AI agents execute much faster than humans. Lee Han-wool's solution is tiered inspection. The first layer detects whether malicious instructions are hidden in external data or tool descriptions, i.e., indirect prompt injection; the second layer directly blocks explicitly prohibited operations; only when a request falls into a gray area does it trigger the third layer, calling the full execution history and task context for deep judgment. This way, most requests take the fast track, and only a few suspicious requests pay the cost of analysis.

Sources:바이라인네트워크

Research

Security Firm Aikido Security Discloses 'wfx-loot' Attack: Over 4,800 Compromised GitHub Accounts Used to Steal Cloud and AI Credentials

Material
2026-11-16 08:00 GMT+8

Cybersecurity firm Aikido Security disclosed a large-scale GitHub credential theft campaign codenamed "wfx-loot" on November 12. The activity began on November 10, with attackers using stolen GitHub access tokens to compromise 4,837 accounts and probe write permissions across 61,764 code repositories. According to Aikido's statistics, at least 1,835 repositories executed malicious programs.

Unlike previous incidents, this attack did not exploit software vulnerabilities in GitHub Actions itself. Instead, attackers impersonated compromised accounts to create pull requests, confirming write access before modifying GitHub Actions workflows or injecting malicious shell scripts. Because some compromised accounts held collaborator or organization-level access across multiple projects, the attack expanded laterally to other development teams' repositories.

When the modified workflows ran, the malware collected sensitive data from the CI/CD environment, including GitHub tokens, cloud service credentials for AWS, Google Cloud, and Microsoft Azure, as well as API keys for AI services like OpenAI and Anthropic. Researchers also found that the malware could create long-term AWS access keys, modify IAM role trust policies, and add Azure federated identity credentials, potentially allowing attackers to maintain persistent access to victims' cloud resources even after the original GitHub tokens were revoked. There is currently no complete statistic confirming the success rate of these persistence operations.

Another security firm, StepSecurity, noted that attackers quickly reverted repository modifications and deleted workflow execution records after the GitHub Actions workflow started, with the entire cleanup process typically completing within one minute. This means enterprises may not detect traces of intrusion when reviewing code post-facto. Researchers advise companies using GitHub Actions to audit abnormal pull requests and workflow change histories since November 10. If malicious execution is confirmed, organizations must immediately rotate relevant tokens, cloud credentials, and API keys, and check cloud platforms for any anomalous permission additions or modifications.

Sources:iThome

Research

GPT-6 Astra Leads Development of Detailed Proof for Lattice Definition of Standard Model: Physicist Nathaniel Craig's Firsthand Account

Material
2026-11-16 06:00 GMT+8

The Standard Model describes known elementary particles and three fundamental forces, agreeing with experiments with precision across eighteen orders of magnitude, from the low energy of hydrogen’s 21-centimeter radio emission to the high-energy collisions at the Large Hadron Collider. However, it has a foundational gap: no one has provided a non-perturbative definition that does not rely on perturbation expansions.

Lattice field theory is a promising alternative to address this issue—discretizing spacetime into a grid to provide a rigorous non-perturbative definition at finite lattice spacing. For strong interactions, this path has been highly successful, allowing the mass of the proton to be calculated from first principles. But the Standard Model is a chiral theory: the weak force treats left-handed and right-handed fermions differently. Placing left-handed fermions on a lattice, the most natural rules inevitably produce an unwanted right-handed partner—the doubling problem.

Ginsparg and Wilson found residual rules for chiral symmetry on the lattice in 1982; Narayanan and Neuberger solved the doubling problem using overlap fermions; and Lüscher further provided methods to distinguish between left-handed and right-handed overlap fermions. Yet a critical question remained unresolved for over 25 years: how to set the phase of the fermion measure. The phase depends on the values of the Standard Model forces at every lattice point. Lüscher only resolved the phase prescription for simpler theories, such as chiral electrodynamics.

Nathaniel Craig, head of the particle theory group at the University of California, Santa Barbara, and permanent member of its Kavli Institute for Theoretical Physics, proposed a refinement path: representing the force fields on the original grid across a series of finer grids, comparing allowed fermion states on adjacent grids to derive part of the phase correction corresponding to small changes in the force fields; summing these parts yields the phase rule on the original grid. The proof demonstrates that this construction provides the correct correction around every allowed loop, and the response of the correction to single-point changes depends primarily on nearby fields, with different mathematical descriptions yielding the same answer.

Craig recounts that GPT-6 Astra suggested the correct way to carry a given set of force fields onto finer grids during this process, showed that comparisons on finer grids would converge to a definite answer, and led the development of the detailed proof. While Craig’s strategy itself was his own, transforming it into a provable construction—including determining the correct refinement scheme, proving convergence, and writing the complete proof—was primarily accomplished by Astra. External lattice field theory experts Maarten Golterman, Andreas Kronfeld, and Yigal Shamir reviewed Astra’s proof and raised key questions, helping Craig build confidence in the results.

Craig explicitly positions this result as a candidate definition, emphasizing that it is far more an answer in principle than in practice, with substantial work remaining before it can be used for actual lattice simulations. He is currently considering whether this construction satisfies reflection positivity at finite lattice spacing—a key condition linking calculations in lattice models to quantum theories where probabilities are conserved. Additionally, Craig mentions that OpenAI’s recent proof regarding the existence of Yang-Mills theory provides the missing puzzle piece: this proof makes it possible to take the continuum limit starting from overlap fermions on the lattice, thereby proving the existence of the simplest unified Yang-Mills theory containing the gauge fields and fermions of the Standard Model.

Sources:OpenAI

Research

Developer Exploits Plain HTTP Certificate Flaw to Localize Coway Air Purifier; Claude Refused Assistance While ChatGPT Provided Firewall Rules

Quick take
2026-11-15 20:30 GMT+8

Developer Scott J. Goldman discovered a security flaw in his Coway Airmega IconS air purifier: during initial setup, the device downloaded its root certificate over unencrypted plain HTTP from an AWS S3 bucket. By exploiting this vulnerability with mitmproxy to intercept and replace the certificate, he enabled his home server to stand in for Coway's cloud service, allowing the device to run entirely locally within Home Assistant.

In implementing this reverse engineering, Goldman used AI assistants to help write network interception and Bluetooth protocol reconstruction code. According to Tom's Hardware, Anthropic's Claude model triggered cybersecurity safety guardrails and refused to assist in configuring a Raspberry Pi to intercept device traffic, even when Goldman stated it was for testing purposes. In contrast, OpenAI's ChatGPT fulfilled the request, generating the required nftables firewall rules; another unnamed AI model rebuilt the fan's Bluetooth communication protocol in approximately 20 minutes.

The project has been released as an open-source integration called 'coway-local'. It allows users to control the device without relying on Coway's IoCare+ cloud servers, addressing previous issues where community integrations frequently broke due to vendor server instability or policy changes (such as app rebranding and HTML layout updates). Currently, the solution has only been experimentally tested on the IconS (AP-1022K) model, and features such as PM10 air quality metrics, filter alerts, and firmware updates are not yet implemented.

Sources:Tom's Hardware

Research

Anthropic Says Claude Fable 5.2 'Paints' Images via Code-Simulated Brushstrokes, Leads Internal Benchmark Against OpenAI Model

Material
2026-11-15 23:42 GMT+8

Developers Jake Eaton and Michael Graczyk at US AI company Anthropic demonstrated on X a unique image generation approach for their new language model, Claude Fable 5.2: instead of relying on traditional diffusion models to compute pixels, it writes Python code to simulate hundreds of thousands of physical brushstrokes on a virtual canvas. According to Eaton, each painting is driven by 1,000 to 2,000 lines of code that model bristle texture, paint application, and error correction. For instance, a photorealistic parking garage scene used approximately 700,000 simulated strokes, while a bird painting was built from 111 coordinate sequences.

This code-based spatial reasoning ability has been shown to transfer to physical robot tasks. AI researcher Aris Zhu tested Fable 5.2 against OpenAI's GPT-6 Astra in controlling a robotic arm to paint Van Gogh's "The Starry Night." Anthropic reported that Fable 5.2 scored 64.9 percent on its internal RoboDojo Sim benchmark, significantly higher than GPT-6 Astra's 29.8 percent. However, Zhu's real-world observations noted that Astra painted faster and was currently the only model able to perform fine motor control tasks like standing up a domino, indicating limitations in Fable 5.2's real-time responsiveness and precision.

Sources:The Decoder

Research

US Judge Denies Motion to Dismiss DRAM Price-Fixing Lawsuit Against Micron, Samsung, and SK Hynix

Material
2026-11-15 23:09 GMT+8

US District Judge Noël Wise in the Northern District of California has denied a motion to dismiss a class-action lawsuit accusing Micron Technology, Samsung Electronics, and SK Hynix of manipulating prices for dynamic random-access memory (DRAM). This ruling allows the case to proceed to discovery and trial phases.

The plaintiffs, comprising 17 individuals and small businesses, allege that the three dominant memory chip manufacturers engaged in parallel conduct—including simultaneous production cuts, a coordinated pivot to high-bandwidth memory (HBM), and the winding down of DDR4 production—to restrict supply and drive up prices. According to Tom's Hardware, DRAM prices surged by as much as 500% in one year amid soaring demand from AI data centers.

The defendants argued that their actions were independent business decisions aimed at maximizing revenue in a highly concentrated market and responding to similar market pressures. However, Judge Wise noted in her decision that the alleged behavior was "more akin to hawks flying in formation—an activity that hawks ordinarily do not do," suggesting that the parallel moves went beyond normal competitive monitoring and constituted plausible evidence of collusion.

The lawsuit also references the companies' involvement in previous price-fixing scandals. While this is not a final judgment or finding of guilt, the court's decision to let the case move forward indicates that the plaintiffs have provided sufficient initial evidence to support their claims. The defendants are required to respond to the original complaint by December 7, with a case management conference scheduled for January 2027.

For investors focused on AI infrastructure, this legal development adds regulatory uncertainty for major memory suppliers. If price manipulation is ultimately proven, these companies could face significant fines or operational restrictions, impacting their profitability and pricing structures across the semiconductor industry.

Sources:Tom's Hardware

Research

You’re all caught up in this view

Sign in / Sign up
ReadingResearchRadarInvestment framework

The company has integrated this mechanism into a platform named Starfort. Among them, Guardian is a real-time guardrail responsible for prompt injection detection, sensitive information leakage interception, and policy violation checks; Opticon is responsible for recording and tracking execution flows; Stratum is responsible for classifying and reviewing outgoing documents. Additionally, there is a Kill Switch function that can cut off AI-related communications by project or organizational unit. Lee Han-wool also mentioned that the company simultaneously develops an attack tool called Stinger for red team testing internally, and the discovered attack patterns are directly fed back into Guardian's defense strategies.

An counter-intuitive argument made during the interview is worth highlighting separately. When the number of agents deployed by an enterprise increases, the volume of operational requests requiring manual approval will surge. Lee Han-wool points out that once approval requests exceed human processing capacity, people will start approving without looking at the background or impact, making humans the new vulnerability themselves. His suggestion is: enterprises must first define which tasks are truly important and require human confirmation, and which can be left to agents for autonomous execution; the focus of approval process design lies not in having someone ultimately confirm the action itself, but in clarifying what tasks, at what timing, and by whom they should be approved, and ensuring that approvers can see sufficient information needed to make judgments.