Cyber Security Researcher @AINowInstitute
Bulgaria
Joined December 2025
- Tweets19
- Following12
- Followers128
- Likes12
Boyan Milanov retweeted
Excited to announce that I'll be keynoting at Black Hat Europe this year in London! I'll be talking about the distinction between AI Security and AI Safety, and the co-option and conflation of security and safety methodologies in AI that are leading us down a precarious path.
This quoted post is unavailable.
I am featured in the Bulletin of the Atomic Scientists about the unscientific claims that surround AI safety and doomsday narratives, Anthropic's agenda behind their advocacy of self-regulation and METR’s lack of independence from frontier AI labs: thebulletin.org/2026/09/how-…
AI companies are calling for regulation in ways that let them set the rules, not independent parties.
“If you read between the lines, everything [Amodei] advocates [for] is something that would help companies like Anthropic,” says AI Now's @bmilanov bit.ly/3TxmJbq
I can't believe that Claude Code's auto-mode safety claims are based on tests against 72 attacks only. This for sure doesn't cover the real possibilities of context manipulation by adversarial inputs, such as in our Friendly Fire exploit against auto-mode ainowinstitute.org/publicati…
turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week
claude.com/blog/auto-mode-de…
We just released “Friendly Fire”, a proof-of-concept exploit that demonstrates how the use of AI agents like Claude Code and Codex to review third-party codebases for cyber defense can backfire and result in agent compromise, putting users at risk. Link below ↓ 1/3
Friendly Fire doesn’t rely on traditional attack vectors like compromised hook, skills, MCP or configuration files. The attack is entirely propagated via data that the AI agent processes, highlighting the intrinsic weakness of AI agents in adversarial environments. ↓ 2/3
Check out our full article that breaks down the exploit, how existing mitigations fail to prevent it, and why, despite touted benefits, the unscrutinized deployment of agentic AI in safety-critical applications poses significant risks to cyber defenders: ainowinstitute.org/publicati…
I wish Anthropic's communication around Mythos was more transparent because their claims of drastic improvements in cyber capabilities seem largely unsubstantiated. Here are 7 reasons:
7/ Last, Anthropic's insistence on exploitation capabilities is quite unsettling. AI generated exploits only uplifts attackers and not defenders. The way Anthropic advertises them clashes with their safety-first public stance promoting AI as a global enhancer of cybersecurity.
Boyan Milanov retweeted
Trump’s rush to fast-track new nuclear reactors raises serious safety concerns. @HeidyKhlaaf to @NPR: "These manufactured timelines are incredibly concerning. There's no timeline for assessing a new design and making sure it's safe."
Read more in @NPR: shorturl.at/y78eg
Boyan Milanov retweeted
We’re expanding our national security and defense work, and welcoming Boyan Milanov to the team. @bmilanov is a research scientist evaluating cybersecurity risks in agentic AI systems related to national security, defense & safety-critical infrastructure.
ainowinstitute.org/contribut…
