Google Launches Gemini 3.5 Flash Cyber AI to Find and Fix Software Vulnerabilities 55%

By Ravie Lakshmanan12%

7/21/2026, 3:09:00 PM

BS Summary: This article contains 21 faulty reasoning types, including Optimism Bias, Attempt to Sell a Product or Service, and Framing Effect, with Appeal to Authority as the most egregious example at 34.1% saturation with 136 hits. Analysis detected 1,086 faulty-reasoning hits from 399 analyzed words, generating a BS Score of 53.2% and a BS Rank of 55% (8,993 of 19,729 articles). This article is worse (more manipulative) than 54.40% of the article peer group.

Google's DeepMind on Tuesday announced the release of Gemini 3.5 Flash Cyber, a specialized artificial intelligence (AI) model built atop 3.5 Flash that's designed to discover, validate, and patch vulnerabilities quickly and efficiently. 
According to the tech giant, the model will be exclusively available to governments and trusted partners via CodeMender as part of a limited-access pilot program. 
CodeMender is an AI-powered agent for vulnerability discovery and patching that was unveiled by the company in October 2025. 
A Google DeepMind spokesperson told The Hacker News that there are plans to extend the model's capabilities to include red-teaming features and end-to-end enterprise defense. 
The lightweight model, per DeepMind, is both cost-efficient and highly capable alternative to large, costly cybersecurity-focused models. 
CodeMender can call upon 3.5 Flash Cyber "multiple times at high speed and low cost," allowing the AI agent to scan more code paths and find vulnerabilities. 
The release of 3.5 Flash Cyber comes alongside Gemini 3.6 Flash and 3.5 Flash-Lite, which are optimized for improved coding, knowledge work, and multimodal performance and low-latency tasks, respectively. 
"Given the dual-use nature of this technology, we have taken an intentional approach to how we deploy 3.5 Flash Cyber," Raluca Ada Popa, DeepMind's Gemini Security Lead, and Four Flynn, vice president of security and privacy at DeepMind, said in a blog post shared with The Hacker News ahead of publication. 
"As part of a limited-access pilot program, 3.5 Flash Cyber will be exclusively available to governments and trusted partners via CodeMender, expanding over time. 
This will give frontline defenders a head start in finding and fixing critical vulnerabilities before they can be exploited, while mitigating against broader misuse." 
Like in the case of Anthropic and OpenAI, Google has put 3.5 Flash Cyber to the test to uncover remote code execution vulnerabilities in public APIs and a memory-corruption vulnerability in a sensitive production service. 
The model is also said to have produced a 100% reliable remote-code execution exploit that bypassed standard mitigation techniques like Address Space Layout Randomization (ASLR) and Write XOR Execute (W^X). 
Google said it's separately bringing CodeMender's foundational capabilities directly to customers with generally available Gemini models through the Gemini Enterprise Agent Platform. 
"By powering CodeMender with 3.5 Flash Cyber, we're providing a highly capable, scalable, and affordable architecture designed to help more defenders secure software," it added. 
Confirmation Bias
6%
Anchoring Bias
0%
Availability Heuristic
6.8%
Representativeness Heuristic
0%
Hindsight Bias
0%
Overconfidence Bias
11.8%
Framing Effect
23.3%
Loss Aversion
0%
Status Quo Bias
6.3%
Sunk Cost Effect
0%
Optimism Bias
31.1%
Pessimism Bias
6%
Negativity Bias
7.5%
Self-Serving Bias
0%
Fundamental Attribution Error
0%
Actor-Observer Bias
0%
In-Group Bias
0%
Out-Group Homogeneity Bias
0%
Halo Effect
10.5%
Horn Effect
0%
Dunning-Kruger Effect
0%
Recency Bias
8.8%
Primacy Effect
0%
Blind-Spot Bias
0%
Ad Hominem
0%
Straw Man
0%
Appeal to Authority
34.1%
False Dilemma
6%
Slippery Slope
0%
Circular Reasoning
0%
Hasty Generalization
11.8%
Red Herring
0%
Bandwagon
0%
Appeal to Emotion
0%
Begging the Question
0%
Post Hoc (False Cause)
6%
Tu Quoque
0%
Burden of Proof
7.5%
Appeal to Nature
0%
Composition/Division
0%
Anecdotal
0%
No True Scotsman
0%
Ambiguity (Equivocation)
3.3%
Gambler’s Fallacy
0%
Middle Ground
0%
Personal Incredulity
0%
Special Pleading
0%
Genetic Fallacy
0%
Unattributed Quote
15.8%
Quote-first Misdirection
19%
Biased Writer Voice
16%
Indoctrination
6%
Politically Left Leaning Bias
0%
Politically Right Leaning Bias
0%
Attempt to Sell a Product or Service
28.6%

399 words analyzed.

Analysis

Hover over highlighted words in the article to view the associated bias or fallacy analysis.