SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Developer finds AI email assistant rates every reply 0.85 confidence, including spam

0
·1 views

A developer building InboxSync, a RAG-based email reply suggestion tool, discovered that the system's confidence score was returning a fixed value of 0.85 for every response regardless of context. Testing revealed the AI confidently drafted replies to spam, out-of-office messages, and even a formal GDPR data deletion request — situations where no reply should be sent at all. In the GDPR case, the system produced an authoritative-sounding legal commitment with unfilled placeholders, which if sent could constitute a binding legal acknowledgment. The root cause was that the confidence metric was never properly wired to any decision logic, making it a meaningless static number rather than a genuine quality signal. The findings highlight a broader risk in AI pipelines: a system that cannot recognize when it is out of its depth may cause real harm by acting confidently in domains it has no training data for.

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 ·

macOS Screen Sharing Flaw CVE-2026-65400 Lets Attackers Gain Root, Install Crypto Miner

A critical authentication bypass vulnerability, CVE-2026-65400, has been discovered in macOS Screen Sharing, affecting macOS Tahoe before 26.6.1, Sequoia before 15.7.9, and Sonoma before 14.8.9. Attackers can connect remotely via TCP port 5900 and exploit an authentication state management flaw to establish a session without valid credentials, requiring no user interaction. In multiple incidents reported to NCSC-NL, threat actors leveraged this access to gain root privileges and install a Monero cryptocurrency miner. Apple has released patched versions of all three affected macOS branches to address the vulnerability. Users and administrators are advised to apply updates immediately, disable unnecessary Screen Sharing, and block TCP/5900 from public internet access.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

GeoServer Zero-Day SQL Injection Draws Mass Probes Within Hours of Disclosure

A critical SQL injection vulnerability in GeoServer's jsonArrayContains filter was publicly disclosed on August 14, 2026, allowing attackers to inject malicious arguments into backend database queries on PostGIS, Oracle JDBC, and H2 data stores. Security firm WatchTowr observed hundreds of exploit attempts from a small number of source IPs within hours of the public disclosure. The flaw can potentially escalate to remote code execution depending on the server's database configuration, though RCE has not been confirmed across all affected deployments. No patches were available at the time of publication, and no successful compromises have been publicly confirmed despite the volume of probing activity. Administrators are advised to isolate GeoServer from public internet access, apply WAF rules to block malformed filter inputs, and enforce database least-privilege principles until an official fix is released.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AmnesiaStealer Malware Uses Fake GitHub Pages to Hijack macOS Browser Sessions

Jamf Threat Labs has identified AmnesiaStealer, a multi-stage Rust-based infostealer targeting macOS users, published in a report dated August 13, 2026. The malware spreads through a fake GitHub ClickFix page that tricks users into pasting a Terminal command, which downloads and executes a malicious shell script. Once installed, it steals login passwords, Keychain data, Apple Notes, Telegram content, Chromium browser profiles, and cryptocurrency wallet information, sending it all to a remote command-and-control server. A secondary module called stream_module uses Chrome DevTools Protocol to silently take over live browser sessions, enabling remote control of tabs, cookies, mouse, and keyboard input. The malware also attempts to establish persistence by creating a root LaunchDaemon, though its exploit of CVE-2020-9771 for TCC bypass was found to fail on current macOS 26 without Full Disk Access.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Critical SAP Commerce Cloud Flaw Under Active Exploit Three Days After Patch Release

A critical unauthenticated remote code execution vulnerability, CVE-2026-58231, affecting SAP Commerce Cloud's Data Hub Adapter was patched by SAP on August 14, 2026. Attackers began attempting to exploit the flaw just three days after the patch was released, with activity detected by the Defused honeypot. The vulnerability allows external attackers to abuse a default authentication client and bypass input validation in the Data Hub Adapter, potentially enabling arbitrary code execution without any user interaction. SAP has issued Security Note 3771065 with a fix, and administrators are urged to update, restrict Data Hub Adapter access to trusted IPs, and isolate the component from the internet. Over 4,200 IPs flagged by Shadowserver represent product fingerprints rather than confirmed compromised or unpatched systems.

0 comments Read more at DEV Community