“Anthropic recently said one of its models found a weakness in the HAWK cryptographic scheme after about 60 hours, even though the system had already gone through roughly two years of expert review." (Instagram carousel, @innovation, sponsored by BTQ Technologies. Slide text: "ANTHROPIC'S AI FOUND WHAT EXPERTS MISSED" / "It was a real…”
Plain restatementAnthropic publicly stated that its Claude Mythos Preview model produced an improved attack against the HAWK post-quantum digital signature scheme in approximately 60 hours of work, after HAWK had already undergone about two years of expert human cryptanalytic review.
This one is essentially accurate. Anthropic really did publish research on July 28, 2026 saying its unreleased Claude Mythos Preview model produced an improved attack on HAWK, a candidate post-quantum signature scheme, in about 60 hours, after HAWK had gone through roughly two years of expert review. The result was real enough that HAWK's own designers withdrew it from the NIST standardization process the next day. Two caveats matter. First, Anthropic describes the HAWK work as a human researcher working together with the model over a week, not the model working alone, and the earlier mathematical groundwork had been laid by human cryptographers, with other researchers and even a different AI model attacking HAWK around the same time. Second, HAWK was never deployed anywhere, and Anthropic states plainly that no production software needs to change, so nothing here breaks the encryption on your bank account or messages. Also worth knowing: the Instagram post is a paid advertisement for BTQ Technologies and its QCIM product, which has no connection to the Anthropic research it is using as a hook.
Anthropic [drifted from the evidence:] recently said one of its [drifted from the evidence:] models found a weakness in the HAWK [drifted from the evidence:] cryptographic scheme [drifted from the evidence:] after about 60 hours, [drifted from the evidence:] even though the system had already [drifted from the evidence:] gone through roughly two years of expert [drifted from the evidence:] review." (Instagram carousel, @innovation, sponsored by BTQ Technologies. Slide text: "ANTHROPIC'S AI FOUND WHAT EXPERTS MISSED" / "It was a real example of AI uncovering a problem that human [drifted from the evidence:] researchers had not found first.")
Anthropic [added by the neutral restatement:] publicly stated that its [added by the neutral restatement:] Claude Mythos Preview model produced an improved attack against the HAWK [added by the neutral restatement:] post-quantum digital signature scheme [added by the neutral restatement:] in approximately 60 hours [added by the neutral restatement:] of work, after HAWK had already [added by the neutral restatement:] undergone about two years of expert human [added by the neutral restatement:] cryptanalytic review.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Anthropic did publicly make this claim. The wording in the Instagram caption tracks Anthropic's own language closely.
- The model name "Claude Mythos Preview" is real and correctly attributed.
- The "about 60 hours" figure is Anthropic's own stated figure.
- The "roughly two years of expert review" framing is also Anthropic's own, referring to two rounds of the NIST additional signatures process.
- The finding was real, independently checked at the reduction level by a cryptographer active in the NIST process, and consequential: HAWK was withdrawn from standardization the following day.
- The caption's own hedge is correct and matches Anthropic: this does not mean AI can break bank accounts, passwords or encrypted messages today.
- Agency overstatement (omitted qualifier). The claim and the slide headline present the model as the finder. Anthropic's own description is that an Anthropic researcher worked together with Claude to develop the HAWK attack over the course of a week. The fully autonomous result was the AES one, not HAWK. The Instagram framing merges the two.
- "Found what experts missed" / "human researchers had not found first" (exaggeration and omitted context). This is only partly right. The reduction path had already been mapped by van Gent and Pulles; human cryptographers had an active HAWK key-recovery preprint in June 2026; and a GPT-5.6-assisted attack on HAWK was posted eleven days before Anthropic's. Matthew Green noted that AI was not the only thing breaking HAWK and that other researchers had recently produced a similar result, while acknowledging Anthropic's was better. The picture is a crowded research area, not a lone AI discovery in a vacuum.
- Scope compression. "Found a weakness in the HAWK cryptographic scheme" omits that the demonstrated end-to-end key recovery applies to HAWK-256, a challenge parameter set, not the deployed-grade parameter sets, and that HAWK was never in production anywhere.
- Marketing-as-evidence (the surrounding post, not the claim itself). The carousel is a paid BTQ Technologies promotion carrying #btqpartner. It uses a genuine, unrelated Anthropic research result and a New York Times headline as an implicit endorsement of BTQ's QCIM product. Nothing in the Anthropic research, the NIST process, or the NYT coverage concerns BTQ or QCIM. The visual pairing invites readers to infer a connection that does not exist.
- Headline framing. "AI is putting the world's digital locks on borrowed time" runs against Anthropic's explicit statement that no production software needs to change and against expert commentary that this was not a step change of the "AI breaks Curve25519" variety.
- The caption's line that "Anthropic says its models could produce new cryptographic ideas faster than researchers could verify them" could not be matched to a specific Anthropic statement. The verification-bottleneck argument is prominently made by Matthew Green in independent commentary, and reporting notes validation took weeks. Attribution of that specific framing to Anthropic is unconfirmed.
- Whether any fully independent party has reproduced the complete end-to-end HAWK-256 key recovery, as opposed to confirming the mathematical reduction, was not established in the sources retrieved.
- The precise share of intellectual contribution between the human researcher and the model in the HAWK result is not externally auditable. Only Anthropic's account of its own process is available.
- The exact 60-hour accounting, whether wall-clock, model-hours, or compute-hours, is not broken out in the sources retrieved.
Anthropic published this research on July 28, 2026. In its own post, Anthropic states that despite HAWK having survived two rounds of expert human review over a period of two years, Mythos was able to improve the best-known attack on it in roughly 60 hours of work, effectively cutting its key strength in half. Anthropic's official announcement thread repeats that HAWK survived two years of expert review and that in 60 hours Mythos Preview found a previously unknown attack that reduced the scheme's key strength by half. The attack exploits a previously unidentified automorphism, a symmetry in the lattice underlying HAWK. Earlier work by Daniël van Gent and Ludo Pulles had established that such a symmetry, if it existed, would reduce HAWK key recovery to a short-vector problem in a lattice of roughly half the original dimension. Their work opened the attack path but did not itself break HAWK. Anthropic says Mythos Preview found the specific automorphism needed to exploit that path. Anthropic explicitly states that neither result has practical impact on today's computer systems and that no production software must change. HAWK was a candidate scheme, not a deployed one. The demonstrated end-to-end key recovery targets HAWK-256, a challenge parameter set deliberately provided as a cryptanalytic target, not the security-level parameter sets HAWK-512 and HAWK-1024. Anthropic followed responsible disclosure: it notified the HAWK authors in June 2026 and coordinated public disclosure to the NIST pqc-forum on July 28. Cryptographer Daniel Apon publicly confirmed the reduction within hours. On July 29, 2026, the HAWK team withdrew HAWK from the NIST additional signature standardization process, saying that mitigations such as doubling parameters would make the scheme uncompetitive. NIST subsequently marked HAWK as withdrawn.
Complete reasoning
The reply receipt is formatted for pasting into the thread where the claim is circulating.
Compact share page: verify.trueseeker.com/s/fb6ca33eeceb/1zyoC4MDQB0CH-BKTeC689L