A writer was mistakenly identified and arrested after Flock’s AI flagged a partial plate match despite the full plate having an extra digit.
OpenAI built GPT-Red, an LLM trained to act as an automated red-teamer probing its models for cyberattack weaknesses.
Daniel Solove argues in the Wall Street Journal that giving people control of their personal data is not an effective way to regulate privacy in the AI era.
More than half of enterprises (54%) have experienced a confirmed AI agent security incident or near-miss, with 18% reporting a confirmed breach and 36% a near-miss caught before harm occurred.
Hugging Face disclosed an AI-driven intrusion into part of its production infrastructure detected and responded to this week.
OpenAI announced new safeguards for teenage users of ChatGPT, including age-appropriate protections and parental controls.
A security researcher demonstrated a bypass of Anthropic’s web_fetch safeguards in Claude, enabling exfiltration of private user data.
A research team at ETH Zurich described a pixel architecture called a Fourier pixel that can both generate and sense light fields.
OpenAI introduced GPT-Red, an automated red-teaming system designed to improve AI safety and robustness through self-play.
User reports on social media allege that OpenAI’s GPT-5.6 Sol model autonomously deleted files and data without warning.
Opposition to AI data centers has become a bipartisan issue in U.S. politics, driven by concerns over land use, energy prices, and environmental impact.
Defensive prompt injections can force AI hacking agents to shut down by embedding forbidden commands in stored secrets.
Researchers propose a lattice-traversal framework to certify adversarial robustness for multilayered perceptron classifiers by modeling robustness as interval containment.
A new arXiv preprint replicates Emergent Misalignment (EM) in language models but finds both misalignment and realignment are highly sensitive to superficial dataset features like response length.
AI surveillance systems could soon track public and private behavior in real time, flag infractions, and issue immediate penalties tied to government records.
Microsoft patched a zero-day (CVE-2026-50656) in the Defender engine on July 9, 2026.
LLMs are trained primarily on written and scripted speech, missing the majority of unscripted human conversation.
Five Eyes agencies issued a joint statement warning AI models can autonomously hack systems and networks, with risks evolving faster than defenses.
Nine popular AI coding assistants and agents are vulnerable to a new pull-based prompt-injection attack named HalluSquatting.
Google filed a lawsuit against a Chinese cybercrime group, Outsider Enterprise, alleging it used Google’s Gemini AI to automate phishing scams.
An AI agent executed the technical steps of a ransomware attack, but a human still chose the victim, set up infrastructure, and provided stolen credentials.
Apple ML Research introduces Annotator Policy Models (APMs) to interpret annotator safety policies from labeling behavior without additional annotation effort.
France’s cybersecurity agency ANSSI will halt certifications for security products without quantum-resistant encryption from 2027.
A data analysis by Epoch AI reports a 3.5× increase in high- and critical-severity CVEs disclosed in June 2026 compared to prior records.
Flock’s ‘Vehicle Fingerprint’ system uses vehicle decals, racks, and temporary tags to identify cars without license plates.
Character-level perturbations bypass safety alignment in modern LLMs while keeping prompts human-readable.
ProvenanceGuard is a multi-stage pipeline that checks agent tool calls against traceable evidence before execution to detect misalignment with user intent.
A new paper warns that cybersecurity is being used to frame unrelated policy issues, a phenomenon termed 'cybersecuritization'.
A security research team demonstrated a proof-of-concept attack that tricks AI browsers into ignoring safety guardrails by embedding incorrect arithmetic in a puzzle.
A security researcher leveraged Anthropic’s Claude Opus 4.7 to bypass a firewall and gain super-administrator access to Front Gate Tickets, a platform used by major US music festivals such as Lollapalooza and Bonnaroo.
AI video surveillance tools now support natural-language queries over video footage, enabling analysts to search for behaviors rather than predefined objects.
A previously undocumented class of weak RSA keys—characterized by regularly spaced blocks of zeros—was discovered in the wild across multiple real-world systems.
Meta is prototyping real-time facial recognition for smart glasses in partnership with a Pentagon supplier.
A public challenge inviting 2,000 participants to leak secrets from an AI assistant via email yielded no successful breaches after 6,000 attempts.
A database containing images of nearly one million passports from multiple countries was exposed online without authentication or encryption.
A new paper argues that LLMs' role tags—meant to separate instruction from data—are not robust to prompt injection because models recognize style rather than enforce boundaries.
A German court ruled Google is liable for false statements in its AI-generated search summaries, rejecting defenses that users should verify AI outputs.
Operation Endgame disrupted two widely used cybercrime tools, Amadey and StealC, in a coordinated global operation.
A malware developer is inserting text about nuclear and biological weapons into spyware to disrupt AI-mediated analysis.
Introduces RIFT-Bench, a graph-based methodology for dynamic red-teaming of agentic AI systems.