SShortSingh.
Back to feed

Developer builds 12-state tracking system to hold AI agents accountable for promises

0
·2 views

A developer has built COGEXT, an accountability layer designed to track commitments made by AI agents through 12 distinct states, from detection to fulfillment or failure. The system addresses a core weakness in current AI agents: they generate intent readily but have no mechanism to track whether those commitments are ever fulfilled. Every extracted commitment is stored in PostgreSQL and transitions between states via a single atomic function, ensuring the audit log always matches the recorded status. A two-stage pipeline first classifies agent output to distinguish genuine commitments from suggestions or hypotheticals, then uses Meta's Llama 3.3-70B model to extract structured fields like action, recipient, and deadline. Fulfillment is verified through a weighted field-coverage model rather than simple keyword matching, requiring evidence to cross a confidence threshold before a commitment is marked complete.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Developer Builds Node.js Workaround for eBay's Missing Sold Listings API

eBay's public Browse API only returns active listings, leaving developers without access to actual transaction prices paid by buyers. The author, who owns a third-party service called CompSniper, demonstrates how to retrieve completed sold listings using a single GET request in Node.js 20 or newer. The tutorial covers filtering by marketplace, item condition, and price range across eight supported eBay regional sites including the US, UK, and Germany. It also addresses error handling, distinguishing between temporary rate limits and quota exhaustion to avoid unnecessary retry loops. A free tier offering 100 monthly requests is available for developers who want to test the integration against live sold-listing data.

0
ProgrammingDEV Community ·

OpenAI's GPT-6 Astra Rated 'Critical' for Cyber Risk as Prompt Injection Threat Persists

OpenAI launched GPT-6 Astra on September 3, 2026, marking it as the first of its models to reach the 'Critical' cybersecurity capability level under the company's own Preparedness Framework, meaning it can autonomously discover and exploit unknown security flaws across hardened systems. The model's system card, citing external evaluations, reports a prompt-injection attack success rate of 8.5%, down from 27% for its predecessor Sol, though security researchers note that improvement is not the same as elimination of the threat. At scale, even an 8.5% success rate across thousands of daily tool-using agent sessions — where attacker-controlled text can enter via uploaded files, emails, or third-party APIs — translates into a meaningful real-world risk. Third-party evaluator Artificial Analysis ranks Astra 14th among 202 tracked models with an Intelligence Index of 60, offering a more measured view than OpenAI's own benchmark figures. Developers building agentic systems are being urged to treat prompt injection as a structural architecture problem rather than a model-level issue, since no lab has yet published a complete fix for the vulnerability class.

0
ProgrammingHacker News ·

Study of 17,000 AI Coding Agent Runs Reveals Tool Preferences of Claude, Codex and Cursor

A study conducted by Armature Tech analyzed approximately 17,000 runs across three major AI coding agents — Claude, Codex, and Cursor — to determine which tools each system prefers to install and use. The research aimed to provide empirical data on the real-world behavior of these agents during coding tasks. Findings were published on the Armature Tech blog, offering developers insight into the default tendencies of popular AI coding tools. Such data can help engineering teams make more informed decisions when choosing or configuring AI-assisted development environments.

0
ProgrammingDEV Community ·

Developer releases fetch-sentinel alpha to block prompt injection in AI web agents

A developer has published fetch-sentinel v0.1-alpha, an open-source, CPU-only local firewall designed to protect AI agent context windows from indirect prompt injection attacks during web fetches. The tool intercepts external web content before it reaches a large language model's context, separating data from potential malicious instructions using four mandatory layers including isolated fetching, structural detection, privilege separation, and signed traceability. The project is backed by 161 passing tests and uses only standard library dependencies plus two lightweight packages, with no native dependencies. The release is explicitly presented as an alpha repository, not production-ready, as an independent audit identified two unresolved structural vulnerabilities — KI-10 and KI-11 — that require a major refactor before the codebase can be considered stable. The tool is licensed under AGPL-3.0 and targets a known attack surface where adversaries embed hidden instructions in HTML comments, Unicode codepoints, or manipulated metadata to hijack autonomous AI agents.