Developer Tests Whether AI Agent Authorization Actually Binds to Final Execution

·1 views

A developer building AgentGuard, an AI agent security tool, investigated whether authorization decisions genuinely bind the specific action that gets executed, rather than merely being associated with it. The core concern is that an authorized action — such as a refund of $100 — could be replaced at execution time with a different one, such as a refund of $1,000, without triggering a rejection. Independent adversarial testing by OpenWorkProof uncovered two high-severity vulnerabilities: one where authorization could be linked to inputs other than the exact execution parameters, and another where an arbitrary callable could be injected into the execution path. In response, the developer redesigned the system around immutable execution snapshots and capability-bound execution to prevent any material change between authorization and actual execution. The findings highlight a broader gap in AI agent security, where permission systems may approve actions without guaranteeing that the approved action is precisely what runs.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Tencent's T-Mem System Anticipates Future Context to Retrieve AI Memories

Researchers at Tencent PCG have developed T-Mem, a long-term memory system for AI agents that addresses a core limitation of existing solutions: the reliance on semantic similarity between stored memories and incoming queries. Instead of searching for similarity at retrieval time, T-Mem pre-saves contextual trigger scenarios at the moment a memory is written, enabling recall even when the current conversation shares no lexical or semantic overlap with the stored information. The system organizes memory across four quadrants combining retrieval direction and memory granularity, each with its own trigger mechanism. T-Mem achieved state-of-the-art scores on two benchmarks — 80.26% on LoCoMo and 74.81% on LoCoMo-Plus — with only a 5.45 percentage point drop on the harder benchmark where mainstream systems decline by 28–50 points. The work has been accepted to the EMNLP 2026 Main Conference and is already deployed in Tencent's QQ AI Partner project.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DeepCybo's PhysBrain 1.5 Tops Open-Source Physical AI Rankings with 72.5 Score

Chinese AI firm DeepCybo released PhysBrain 1.5 on September 9, 2026, a physical foundation model designed to bridge the gap between language AI and real-world robotic action. The model achieved an average score of 72.5 across 28 public benchmarks, claiming the top spot among open-source physical AI models and trailing leading closed-source competitors by just under one point. PhysBrain 1.5 is built on a closed-loop architecture called Physical Loop, which enables robots to observe, reason, act, and self-correct based on environmental feedback. DeepCybo trained the model using Ego360, a proprietary human-interaction dataset derived from panoramic video capturing full-body movement and task-level voice data, rather than internet text. Both the 2B and 8B variants are released under an Apache 2.0 open-source license and attracted over 3,000 Hugging Face downloads within three days of launch.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AllSpark's Open-Source Iris Search Agent Rivals Closed-Source Tools with Context Management

AllSpark Research has launched Iris, an open-source AI search agent available in two sizes — Iris-mini (35B total, 3B active parameters) and Iris-pro (397B total, 17B active parameters) — both built on a Mixture-of-Experts architecture with 256K context windows and released under the Apache 2.0 license. Iris achieves strong benchmark scores, including 88.6 on BrowseComp and 92.9 on DeepSearchQA for the pro model, competing closely with previously dominant closed-source search agents. A key finding from the research is that intelligent context management — rather than simply scaling up parameters — is the primary driver of performance, with the smaller Iris-mini benefiting more from context strategies than the larger model. The agents were trained using an alternating supervised fine-tuning and reinforcement learning process run against live web search, not static offline data. By releasing model weights, context management strategies, and data construction methods together, AllSpark aims to close a two-year gap between open- and closed-source search agent capabilities.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Fix Emails Landing in Spam: Debugging SPF, DKIM, and DMARC Alignment

Emails landing in spam despite configured authentication often stem from a misalignment between the domain in the visible From header and the domains used by SPF or DKIM. DMARC passes only when at least one authenticated identifier — either the SPF envelope-from domain or the DKIM d= signing domain — organizationally matches the visible From domain. Two separate authentication passes using unrelated domains will still cause DMARC alignment to fail, a common source of confusion during DNS migrations. To diagnose issues accurately, administrators should inspect the full headers of a received message and check the Authentication-Results field rather than relying on setup dashboards. Alignment rules can be relaxed or strict depending on the aspf and adkim tags in the DMARC record, and these settings become especially critical when moving DNS zones or introducing subdomains.

0 comments Read more at DEV Community