Study: ChatGPT Named Only 4 of 48 AI Startups But Recommended 28 by Category

·3 views

A developer tested 48 AI-native startups by asking a language model to identify them by name, and only 4 — ElevenLabs, Suno, Runway, and Cursor — were correctly described. The remaining 44, including well-funded companies like Decagon and Mercor, returned a near-identical response stating the model had no reliable information about them. However, when a search-grounded model was asked category-based questions without brand names, 28 of the 48 products appeared in recommended tool lists. The experiment highlights that AI model name recognition depends heavily on training data, including press coverage, reviews, and third-party mentions that most young startups lack. The author argues founders should focus less on brand recall and more on category visibility, since half the sample was invisible by name yet discoverable by function.

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)

F-RevoCRM XSS Flaw CVE-2026-71368 Puts Logged-In Users at Risk

A medium-severity cross-site scripting vulnerability, tracked as CVE-2026-71368, has been disclosed in F-RevoCRM versions 7.3.0 through 8.0.3, with a fix available in version 8.0.4. The flaw, published by JVN on August 17, 2026, allows an attacker to craft a malicious URL that, when opened by a logged-in user, executes arbitrary JavaScript within the CRM's origin. This can result in unintended CRM operations or the theft of session information using the victim's existing privileges. No authentication is required from the attacker, but the attack does depend on a valid user session and some form of user interaction, such as clicking a link sent via email or chat. Users are advised to update to version 8.0.4 immediately and to avoid opening untrusted links while logged in to the platform.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Microsoft Uses Multi-AI System to Generate Synthetic Attack Logs for Detection Testing

Microsoft's Defender Security Research Team has published a method that uses AI to generate synthetic attack logs based on MITRE ATT&CK techniques, aiming to address the chronic shortage of real-world attack data for building detection rules. The system takes attack tactics, specific procedures, and context such as target OS as inputs, then produces structured logs containing process names, parent-child relationships, command lines, and event sequences. A multi-AI pipeline — where separate models handle generation, evaluation, and improvement in a feedback loop — outperformed single-model approaches, especially for complex multi-step attack scenarios. The research was published on May 12, 2026, and is positioned as a lab-testing aid rather than a replacement for ground-truth logs captured from actual attack reproductions. Microsoft notes that synthetic logs cannot confirm real-world attack reproducibility but can meaningfully accelerate detection engineering where real data is scarce, sensitive, or costly to collect.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CrowdStrike Study Tests AI Accuracy in Auto-Triaging SOC Alerts on Windows Endpoints

CrowdStrike researchers published findings on using AI to automatically classify Windows endpoint security alerts as genuine attacks or false positives, aiming to reduce the burden on human Security Operations Center analysts. The system pairs a classification AI, which reads alert data and provides a judgment with reasoning, with a calibration AI that estimates the confidence level of each decision. High-confidence results are routed for automated processing or priority investigation, while low-confidence ones are escalated to human reviewers. The study found that classification accuracy can degrade over time due to distribution shift caused by new attack patterns or product updates, making continuous monitoring essential. Researchers concluded that AI-only automated alert closure requires ongoing accuracy checks and threshold adjustments to remain reliable in production environments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Unpatched Microsoft Defender flaw lets local attackers escalate privileges to SYSTEM

A critical zero-day vulnerability tracked as CVE-2026-69414, dubbed ShieldBreak, has been disclosed in Microsoft's Malware Protection Engine, affecting Windows 10, Windows 11, and Windows Server. The flaw stems from an incomplete fix for a prior vulnerability, CVE-2026-50656, and allows an attacker who already has low-privilege code execution on a device to escalate to full SYSTEM-level privileges. Microsoft has acknowledged the issue and is working on a patch, but no fix has been released as of August 17, 2026. A public proof-of-concept exploit exists, though active in-the-wild exploitation has not yet been confirmed. Organizations are advised to monitor for abnormal privilege-crossing process activity around Defender services while awaiting an official patch.

0 comments Read more at DEV Community