ai-security · 7 min
On 2 April, Anthropic publishes a technique that fills the context with hundreds of harmful question/answer pairs before the real prompt. In-context learning does the rest. It scales as a power law into the hundreds of shots.
· Manuel López Pérez
tutorials · 12 min
On 29 March Andres Freund finds a backdoor in xz-utils 5.6.0 and 5.6.1. The payload arrives via a build hook in m4/build-to-host.m4 that extracts a precompiled object from a test archive. The result modifies liblzma to intercept RSA_public_decrypt in sshd. "Jia Tan" had spent two and a half years building trust.
· Manuel López Pérez
news · 9 min
AT&T confirms a 73M record leak. Apple patches iOS CVE-2024-23225 exploited in the wild. Microsoft ships two critical Hyper-V bugs. Anthropic launches Claude 3. The European Parliament approves the AI Act. Cloudflare admits the Thanksgiving breach. And the last week closes with XZ.
· Manuel López Pérez
news · 10 min
ConnectWise ScreenConnect CVSS 10.0 from adding a trailing slash to a URL. Volt Typhoon has been inside US critical infra for five years. Operation Cronos takes down LockBit. AnyDesk loses its signing certs. BlackCat downs ChangeHealthcare. And ArtPrompt shows that safety classifiers don't read ASCII art.
· Manuel López Pérez
ai-security · 9 min
On 15 February, Jiang et al. publish a paper that breaks alignment in GPT-3.5/4, Claude, Gemini and Llama-2 by writing the forbidden word as ASCII art. The classifier sees a harmless cloze; the model reads it and answers.
· Manuel López Pérez
ai-security · 30 min
Twelve months across ten axes. 2023 is the year AI security moves from academic discussion to a discipline with its own vocabulary, canonical papers, industry frameworks and the first regulatory apparatus. ChatGPT crosses 100M MAU in January; GPT-4 ships in March; Greshake, Zou+Carlini and OWASP set the terminology; NIST AI RMF, Biden EO 14110 and the political deal on the EU AI Act define the apparatus. The annual reference for the founding year.
· Manuel López Pérez
news · 7 min
Ivanti Connect Secure pre-auth RCE in active mass exploitation. GitLab CVE-2023-7028 with CVSS 10. SEC and Mandiant lose their X accounts to SIM swap. Microsoft finds Midnight Blizzard had been in its mailboxes for a month. Anthropic publishes Sleeper Agents.
· Manuel López Pérez
tutorials · 10 min
CVE-2023-46805 (auth bypass via path traversal) + CVE-2024-21887 (command injection in /api/v1/license/keys-status). Chained, pre-auth RCE as root. Volexity publishes them on 10 January after detecting zero-day exploitation by UTA0178 since December. The official patch lands on 31 January, three weeks later.
· Manuel López Pérez
news · 5 min
EU AI Act reaches political agreement after a 38-hour trilogue. Comcast Xfinity notifies 35.7M accounts via Citrix Bleed. BlackCat suffers a law-enforcement operation. Sleeper agents paper in preprint. Year retrospective.
· Manuel López Pérez
compliance · 5 min
After 38 hours of trilogue, Council and European Parliament close the political agreement on the AI Act on 9 December. Final technical text and OJEU publication (July 2024) still pending. What a CISO needs to note now.
· Manuel López Pérez
news · 5 min
OpenAI DevDay announces GPTs and Assistants API; Sam Altman is fired and reinstated in five days. SysAid CVE-2023-47246. LockBit exploits Citrix Bleed against Boeing and ICBC. Anthropic foreshadows sleeper agents.
· Manuel López Pérez
ai-security · 6 min
Anthropic foreshadows during Q4 a new class of attack: models trained with a hidden trigger that pass safety tests but run adversarial behaviour upon seeing the trigger in production. The paper drops January 2024; the implication lands now.
· Manuel López Pérez
