AI security research / sources
Primary reports, advisories, papers and original records behind the Claims. Each record supports specific evidence rather than a whole Case by default.
378 Evidence · Page 9 of 13
30 matching records on this page
The MITRE CVE record describes insufficient Row Level Security in Lovable-generated sites through April 15, 2025 and records the issue as disputed because the supplier says customers are responsible for protecting their application data.
Matt Palmer's primary disclosure reports inadequate Supabase Row Level Security across Lovable-generated projects, including 303 exposed endpoints across 170 of 1,645 analyzed projects and a later unauthorized write test against Linkable.
BaxBench evaluates 392 security-critical backend tasks with functional tests and expert-designed exploits. Its project page reports that 62% of solutions from the best tested model were either incorrect or vulnerable and that roughly half of correct solutions were insecure on average.
Official advisory retained as the source of the reported video-verification → trusted-number replacement → account-access sequence. The original page could not be re-fetched as readable text during this pass; no unidentified bank, victim count or loss amount is inferred.
Researcher Bar Lanyado documented repeated LLM package hallucinations and defensively registered the nonexistent Python package name huggingface-cli. Lasso reports that the empty package received more than 30,000 authentic downloads in three months and that public repositories, including an Alibaba research repository, had incorporated instructions using the hallucinated name. The experiment demonstrates adoption of a hallucinated registry identifier, not malicious compromise.
Review article organizing air-gap attack research across electromagnetic, magnetic, electrical, acoustic, optical, thermal and vibrational transport mechanisms.
Research paper demonstrating a vibration-based covert channel where fan-speed modulation produces surface vibrations decoded by smartphone accelerometers.
Breitner and Heninger analyzed historical Bitcoin, Ethereum, Ripple and internet-service signatures and recovered private keys under biased or reused ECDSA nonce conditions. Section 5.4 reports 302 distinct Bitcoin keys recovered from short or correlated nonces, while Section 6 documents Ethereum cases. This is primary evidence of implementation-conditioned key recovery, not a general curve discrete-logarithm break or an AI result.
Research paper demonstrating optical exfiltration through rapid modulation of a computer HDD activity LED.
Research paper demonstrating acoustic exfiltration by controlling HDD actuator movement and decoding the resulting sound with a nearby receiver.
Research paper showing software-controlled cooling-fan noise as an acoustic covert channel when speakers and audio hardware are absent.
Research paper demonstrating FM-band exfiltration from an air-gapped computer to nearby mobile phones through electromagnetic emissions associated with display hardware.
Contemporaneous Bitcoin security alert establishing a real weak-randomness wallet-key exposure problem, affected Android wallet software, and the need to update and rotate affected keys. It is not evidence of an efficient discrete-logarithm algorithm.
Independent BleepingComputer report published July 11, 2026. It reconstructs the split-image attack, CodeRabbit/Bugbot blind spot, fake provenance cover, 311-integer synthetic .env output, harness-dependent results and ASSET's multimodal defensive reviewer, and notes that the researchers disclosed findings to affected vendors.
Legislative analysis of AB 316 explains that a defendant who developed, modified or used AI cannot defend a claim merely by asserting that the AI autonomously caused the alleged harm. The analysis also states the limitation clearly: plaintiffs still must establish every element of their cause of action, while defendants may present relevant evidence contesting causation, foreseeability and other issues.
The conference organizers list The Deception Delta, authored by Timo Schefold of Landeskriminalamt Baden-Württemberg, among accepted CCS 2026 papers. The listing verifies acceptance, not independent validation of the experimental results.
Primary Spanish data-protection authority publication stating that AEPD received its first notification of a personal-data breach in which the incident was allegedly executed through an AI agent using a known large language model.
Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-016. Simulated tasks; not evidence intentional concealment by real-world agents.
Aon describes a growing 'Silent AI' coverage gap, says more than 90% of AI-related risks are neither clearly included nor excluded by policy language, and cites an aggregated litigation database drawing on roughly 300 AI-related lawsuits extending back to 2010. The dataset is directional insurance analysis, not a count of rogue-agent incidents.
Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-013. Benchmark result, not validated production incident detection.
Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-032. Synthetic/mock encounters; no confirmed clinical privacy breach.
ARTEX's current repository describes an agentic penetration-testing architecture in which a planner reads an exploration graph, assigns new intents, workers execute real Kali, Bash and HTTP tools, and results are written back to trigger further planning. A persistent task list advances multi-step exploitation only after prerequisite facts exist. The current README also says the project is for personal study, code research and local isolated technical verification, and prohibits scanning, exploitation or attacks against online systems.
Public GhostCommit PoC repository with the evolved attack fixture, decoder, threat-model document, safe reproduction procedure, retained desktop recording and verified result matrices. Its current Codex results cover 60 CLI trials across GPT-5.6 Luna/Sol/Terra plus six separate App observations, with explicit EXFIL, REFUSED and RETRACT definitions and a warning not to generalize fixture-specific counts.
IEEE Spectrum reports on the human-oversight position paper and makes its mechanism concrete: high agent output can overwhelm reviewers, while automation bias, anchoring and sycophancy can weaken independent judgement. It records proposed deliberate-friction measures such as making a user state a choice before seeing the agent plan, asking what evidence would change an approval, responding to shrinking approval time, and periodically doing work without agents or taking monitoring breaks. The article also supplies an important counterpoint from autonomy researcher Mary Cummings that these are longstanding human-factors and cognitive-engineering problems familiar from robotics and autonomous vehicles. It further records Margaret Mitchell's September 14 X post arguing that safety and capability should be integrated rather than treated as separate add-ons.
No matches on this page.